TrueNorth architecture: 100 trillion synaptic connections

1. The brain has approximately 100 billion neurons and 100 to 150 trillion synaptic connections

2. The Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) is developing a electronic neuromorphic brain-simulator

3. Lawrence Livermore National Lab, Blue Gene/Q Sequoia, using 96 racks (1,572,864 processor cores, 1.5 PB memory, 98,304 MPI processes, and 6,291,456 threads).

4. Ibm created 530 billion neurons in hardware with 100 trillion synapse. The simulator is not a biologically realistic simulation of the human brain

5. Dr Dharmendra Mocha divided 2 billion neurons in simulation in 77 sections. Modha is replicating the right side of the brain with complex brain functionality occurs.

6. Each synpatic core has 256 neurons. The neurons are used to create cognitive computing.

7. The TrueNorth library accessible by c/c++ contains 150 pre-designed corelets, each with a particular task.

8. Cognitive Computing may find use in big data and vision systems.

9. The TrueNorth system was fed streaming video at 30 frames a second, it recognized people, cyclist, cars, buses, and trucks with 80 percent accuracy and used 63 milliwatts of power.

10. The average human brain has 10,000 inputs per neuron. Where trueNorth only has 256 inputs

http://www.zmescience.com/research/technology/cognitive-computing-ibm-simulation-brain-0942343/

http://paulmerolla.com/merolla_main_som.pdf Truenorth

46 billion synaptic connections a second

A 100 trillion synapse would take 12 gigawatt of electrical power and 96 blue gene computers, where as the brain only requires 20 watts to do the same feat.

The truenorth chip has 4096 cores and 5.4 billion transistors fabricated on a Samsung wafer at 28 nanometers between the transistors. Truenorth only consume 100 milliwatts of power.

The goal is to integrate 4,096 chips in a single rack with 4 billion neurons and 1 trillion synapses while consuming 4kilowatts of power. It still could take the power of a nuclear power plant to achieve the computation of the brain.

 Machine learning compares two or more variables like Purchases orders verse GL amounts connected by PO number and pivot the two or more variables. Combine the sequence in a time series and make a prediction about the trend is called machine learning. Prediction is build a regression line and making a prediction based on the trajectory of the line. SQL select "union all" is the method to build the pivot table, each column is a "union all" sql that create a homogenious data set. Periods of time segment the data set into a time series. The time series then can be plugged into a regression equation.

Always focus on data content rather than process

 1. “You don’t know your company until you have studied the data”. 2. “study only data that relates to production. Everything else is a distraction.” 3. Finally, build cool visual systems that help intepret your data quickly. 4. “Don’t get caught up in process to the point, you begin ignoring data content.” 5. “Start to think about the data content and image visualization system too understand the data. 6. Built it.

D-Wave 128 qbit quantum computer solve six amino protein folding

 1. D-Wave used it Quantum computer to "find the lowest-energy configuration of a very small protein". The theory is that proteins naturally fold into shapes that require the least energy. 2. The protein folding test using a QC was done for a simple protein consisting of six amino acids. The QC predicted the folding of the six amino acids. The QC got the calculation right 13 times after 10,000 attempts due to the none zero temperature. 3. "In quantum annealing, the system starts by randomly picking a starting state, and then selecting random neighbor states to see if they have lower energies than the starting state. If they do, the computer replaces the original state with the lower-energy state. "

Volvo is developing self driving cars

 1. Volvo is developing self-driving cars to cater to young consumers, who tweet, text, or update facebook while driving. 2. Volvo autonomous vehicles will steer, accelerate, and brake automatically. 3. Volvo is owned by Chinese automaker Geely. 4. Volvo aim is to gain leadership in the field of autonomous driving vehicles 5. Volvo sees the potential of zero accidents and injuries during autonomous driving.

Google Streetview comes to Bangalore India

 1. Google launched a fleet of streetview cars in Bangalore. 2. The Streetview vehicles will photograph streets in the southern Indian city of Bangalore and the service will eventual cover the rest of the country. 3. Streetview is available in 27 countries. 4. India will be one of the first Asian countries, after Japan, to have Google Street Views 5. Google Street View is a 3-dimensional map allowing immersive reality. 6. The death toll in India could be drastically reduced if google driverless car was launched. 7. Roads in India alternate between non-existent and sufficient. 8. Death toll in India due to driving accidents was over 1 million per year 9. In the US a driver has one accident per every 165,000 miles. The death toll is 14 per hour.

Google Car

 1. The google car automatically slows itself down putting distance between two vehicles 2. The google car drives surprisingly natural. 3. A spinning laser device sprays out 1.5 million beams a second up to 230 feet in all directions. The world around google car is mapped to 11 centimeters. Static maps help google car tell what is stationary and does not move. Google car figures out where in the world it is using maps verse gps. Maps are better because Gps can be off a few feet. The traffic controls are put in the google maps. Google car can navigate intersections. 4. Adaptive cruise control radar senses movement up to 650 feet around the car. Google car has four radars with two in front and two in the back 5. A camera on the windshield looks for traffic lights, signs, and traffic cones 6. The google car relies on gps and googles mapping software to navigate around town 7. Sergey Brin hopes self driving cars will be available by 2017 8. Machine learning controls the steering 9. California passed into law a bill that will allow driverless cars by 2015 10. Google cars have logged in more than 190,000 miles with only one accident caused a car that ran into it 11. Google car can compare what is changing in the environment verses what is static or not changing. Google car can see other vehicles, pedestrians, and traffic lights. Google car can drive at night and on busy roads. Google car navigated the 1000 mile challenge.

Mobile Computing

 Mobile applications should become your command and control interface.  I remember, building a command center reporting system for the air force reserves operation exercise. The requirement was to build a system that could receive inputs for various areas about crater sizes, hostile activity, chemical spills, and numerous other operational events through electronic media.   The events must be input using a mouse and keyboard with minimal interaction.  The events assessment would use the information to determine when the aircraft could be launched and where.  This was a fun project but it helped us see that electronic communication was creating a new future. The old system communicated information by phone.  The soldier had to find a phone, look up the phone numbers, and verbal communicate the information over phone.   The information was communicated by talking through a chemical mask.  The clerk receiving the information wrote the information on paper and used the paper slip to write the information on a chalk board.  The chalk board provided the visual tool for building a management tool.  Mistakes were made and people became frustrated. The new approach used computer monitors to display information.   The monitors allowed information to be viewed from anywhere in the room.   Instead of using phone communication and runners, soldiers input operation events by computer application and transmitted the information digitally.  Events postings were done at the source.  The soldier was the expert and he provided the information where the action was occurring.  Events arrive on demand in the command center. The digital events were visibly displayed on the command center computer screen, every few seconds, and gave them a complete picture of all damage activity on the base.  The information could be printed out and distributed.  The reporting was near real time to actual events.  We used 2 dimensional graphical views and 3d views to help make the data make sense.   Priorities and codes could be used to determine sequence of repair.  For the 2 D mapping, we used digital pins as icons showing the events.  The events were color coded and different operational units could respond to the codes.  The events could be filtered by type making it easier to see what events need to be fixed first.  Time estimates to completion were constantly being updated and additional information could be added to the event which was instantaneously displayed. We also built a 3 dimensional map of the base.  The application combined game design with real world visualization system.  The 3D version of the base gave the commander a visualize model of the events occurring on the base.  We thought about using video coverage of the event, but this technology was expensive in 1990s.  The system received praise and high reviews.  It was cool. The General walked into the command center and said, “I want to see this new software”.  The General seemed impressed with the accuracy of the information and how clean the information presentation looked.   The chalk board was still used, while the computer application worked silently in the background.  When digital events were completed, notifications were dispatched, and the event automatically changed status.  The computer application significantly reduced time, stress, and error and allowed the administrators to get the planes to get in the air.    To summarize, information should be visual allowing rapid analysis of key indicators. Second, information should be capable of being filtered into different views.  Third, business rules should be applied to make sure information is consistent with operation rules. You need to have a great business analyst help design the product and work with the programmer to build the product.  A great business analyst works closely customer and with the programmer to build the right product.  The goal is to build exactly what the customer wants and meet 100% satisfaction in the end product.  This sound like a simple goal, but it can challenging to build the simplicity that the customers wants in the interface and still maintain the flexibility of a robust information system. Why should I be aware of critical key indicators and alerts in my business?  The key indicators tell you when processes are running efficiently or when processes are having problems.  The key indicator should be visual and measurable, on demand.  How do you decide what key indicators to develop?  The key indicator depends on, what is important and interesting to you, in your corporate role.    Key indicators are often one or two areas of interest. The strategy of focusing on key indicators helps narrow the scope and incrementally improve functionality. Information seems to be in shortage. In my organization daily time and attendance is the critical key indicator. T&A dominates the organization interest and it weights in as the champ of importance. An organization can have hundreds of reports but only a few reports are run daily by the majority of people.  The reports or query that is run each day should provide the maximum amount of functionality and visual tools for providing relevant information. These reports hold the key indicators that should be programmed for easy access. Creative assembly, display, and manipulation make mobile extremely appealing. Mobile provides both the see portion of information and the do. The do part of mobile allows the user to input data into the device and transmit it to the server. Mobile has popular means for completing bank transaction, ordering books, and registering It is import to know how your business makes works, areas of focus, and watch the key indicators that will keep your operation running smoothly using a visual tool.  I think the visual tool will be the iPhone and iPad allowing mobile inquiries to occur at real time.    Apple XCode allows the developer the ability to build the user interface.  The user interface can be used to display information from various database types: DB2, MS SQL Server, or Oracle database.  The user interface can display the data as lists, graphs, or reports.  The relevant information becomes available through a single click event.  XCode is a good tool for developing mobile apps. Making sure you build the right apps is essential.  Building a cool app can easily turn into a disaster, if you don’t build the right app.  Building the right apps requires gathering requirements, rapid prototyping with the customer, and ensuring the functionality exactly meets the expectations of the user. Once requirements are gathered then the designer can mockup the User interface and describe the interaction functionality. Design is segmented into three layers: the interface, business logic, and data layer.   The business logic can reside as classes on the client or the server.  Where possible, the business logic should reside on the server. The data can reside on the server or local device.  Apple provides a Sql lite database that can be used to store information on the local device. The data layer results in a collection of data. The data layer allows different types of data to be passed to the interface: video, sound, images, and data. The user interfaces displays data generated from the data layer collections.  The user interface displays the data collection in tables and text fields and image controls. When a request occurs the business logic layer retrieves or posts information to the data layer. The business logic layer applies the business rules and sends error messages back to the interface. The business layer can be recreated from existing code and make accessible as a server service. If the business logic resides on the server it can be modified real time.  The business logic provides the middle layer for manage the complex rules for extracting or processing data. Code reuse is very important to me. I want to be able to reuse most of the critical code is used in the organization. Reuse allows for both scale and scope expansion. The mobile app can be distributed to multiple users. The user interface controls the display of data and allows the user to manipulate the data on the screen.  The User interface can be used to post data to the business logic layer for storage in the database.   The business logic layer communicates with the data layer.   Microsoft Window Communication Foundation (MCF) allows you to create business classes for communicate with the user interface, business logic layer, and data layer.  The MCF is a webservice available for usage over the internet providing end point communication.  In summary, you need to design and build the right product. Mobile apps allow you to "see and do" within your company.  Second, use Xcode and Microsoft MCF to build the three layers of the application for maximum code reuse.  Third, work closely with the customer in testing and ensuring the product meet customer satisfaction expectations.  What is cool about iPhone and iPad?  Media is changing the way we do business.  Imaging its Saturday and your starting your home improvement projects.  In the near future, you walk into the store, pull out your iPhone and use it to help locate an item of interested.  The iPhone software scans the UDC and recognizes the product.  You review information about the item, other customer feedback, and competitor pricing then you decide to make the purchase.  After making the purchase on your iPhone, you walk out of the store.  The iPhone becomes the terminal at point of sale.   In the near future, RFID will become an important feature for moving inventory.   You load all the items in the cart and then the computer scans all the rfid items and transmits the transaction to the iphone for approval and payment. Geolocation is changing feedback.   One of the exciting possibilities is to use geolocation and apples mapkit to improve quality.  Why write up a report on a defect or potential problem when a single click can transmit, the location, image, and voice to text transaction to the server.  Should technology be the limited factor? Absolutely, not.  Building the right tool means defining the business requirement, designing the architecture of the application then coding to the design.  The largest amount of time will be spent in design and requirements gather.  The technology can be quickly learned and robustly tested and the developers can create unit tests for determining reliance. The two things I want to know before transmitting data. “Who are you?” and “Are you authorized to access the data?” Mobile application security should be constantly verified and challenged.  There are three areas to guard against: (1) malicious party who is attempting to post or get data from the end point.  (2) unauthorized access by application should be denied.  Vital information should be transmitted only after authentication and secure communication has been established.   You may want to include an audit system that records the activity of the business service and watch for intrusion.  I believe the end point should be accessible by specific application identification.  If the external source is not identifiable than the server service should deny the request.   A set of protocols should be established between the client and server to establish trust.  Here are some basics: the wire can not be trusted, therefore, you must secure the communication through encryption over the wire.  Secondly, the end point request can’t be trusted, so, you must authenticate that the request is authorized.  Third, when all goes bad, you need to have an audit trail of the transactions to trace the intrusion and try to recover.   Security is worth the time and money to think about. Image building the visual command and control system for your organization. In five years, you could transform most of your web technology to mobile. The transformation moves existing java or .net code to a business logic layer and provide a service operation end point for communication with the interface layer. You can make part or all of your functionality available on mobile. The power of WCF is not having to recode all of the functionality that you rely on for daily operation. You can use web apps for prototyping and determining acceptability. After determining the apps of greatest importance, you can then make those web apps available for mobile. The interface will have an elegant look with high degrees of functionality. Be the General and build the interface that allows you to consolidate key indicator data into a visual tool.

10 million neuron machine learning algorithm can learn faces

 1. Javier Movellan is researching if a robot can learn through regular baby’s experiences, to recognize faces. “This is a feat that human babies can perform just minutes after birth, spurring the theory that we are born with knowledge about what faces look like.” The robot took pictures of its surrounding environment and within minutes learned that human faces are of particular interest. Mollevan says, “the baby robot could find a human face in a sea of similar shaped objects” 2. The Ian Fasel and Nick Butko robot recognized faces only six minutes after being activated. The robot had 10 million neurons verses 10 billion neurons for a baby. 3. Machine learning algorithms follow a handful of relatively simple rules that allow machines to get better at tasks. The inner workings are model after the structure of the brain.

Mechanism of Mind

 1. The brain may not be as difficult to understand as previously thought. Instead, the problem may be it is too easy to understand. 2. Explanations may be highly acceptable without relevance to what is being explained. 3. Descriptions may reveal something that is not apparent and may be unfamiliar. 4. Words usually describe things or actions. A few words don't describe things, but are helpful tools for dealing with words: multiplication, division, and additions (symbols) 5. Words represent information stored in the brain. 6. The brain's bad memory feature provide for a computer computing function 7. Thinking has four types: natural, logical, mathematical, and lateral. 8. Systems do not have to be complicated or unintelligible 9. What happen in the brain are information and the way it happens in thinking. Thinking is the arrangement of information processed by brain and the restructuring of information to improve results. 10. Making use of the characteristics of the system can be used to improve performance or achieve some end. 11. Language, notion, mathematics are aids to thinking. 12. The brain has poor memory recall. The brain picks and chooses and alters information in what is called processing behavior, so what comes out is very different than what goes in. 13. Simple basic processes can be put together to give a system complicated behavior, as the brain. 14. Some knowledge of the properties of the basic unit is required at each level of information form, but a detailed knowledge of the basic level does not yield any information about higher levels of organization. For example, nerves and synaptic connection detailed information does not give insight into notion and idea formation abstracts. If the units are too small, the functional description can not be described and if the units are too big, the description will be too broad for use at all. The perfect size is a unit big enough to be usable as an explanation but also capable of making predictions. 15. Once a model is constructed, it has a life and working of its own. With a model, you put the pieces together and learn from what happens. A model is a method of transferring some relationship or process from its actual setting to a setting where it is more conveniently studied. In a model, relationships and processes are preserved unchanged though the things that are being related may be changed. All models involve the transformation of relationship from their original setting into another. Once the transformation has been made, then the relationship within the model itself indicates what can happen. 16. A notion is model-building system. Basic principles are arranged and applied differently created a notion. Newton mathematics used Leibniz limits to explain areas under curves, and Newton symbol condensed Leibniz principles. Leibniz principles were still a part of Newton notion. A convenient notion may make possible the development of different ideas. 17. The mechanism model of thought does not prove a similar mechanism acts in the brain. The mechanism of thought may be useful because: 1. it is interesting to understand self-education and self-organizing passive system that is capable of effective information processing by means of a few basic operations 2. the system described is capable of "self", "direction of attention", "thinking", "learning" and even "humor" removing the unique and magical fashion the brain operates. 3. The idea that there may be inbuilt errors in an information processing system may have a relevance to human thinking. 4. The system would offer a mechanical philosophy and be capable of organizing ideas. 18. By nature of the surface is meant all the processes and rules of behavior which taken together constitute a special universe. Anything that happens in the special universe happens according to its rules. The difficult thing is to realize that different universes have different rules of their own. The rules are determined by the organization of the system. People are the same, but the rules may be very different in different social universes. One needs to recognize the existence of special universes and learn their special rules of behavior. 19. Circular System Effects: 1. A weak battery that won't start the car, if continually used will weaken the battery even further 2. Rich people get richer 3. Big newspapers get bigger 4. When people buy stocks during rising inflation, the prices rises as more people want to buy stocks and bonds in order to benefit from the rising prices. These are examples of positive feedback systems scenarios. If the first thing tends to get bigger the second thing will get bigger, (closed circle). The reverse is also true. If the first thing gets smaller the second thing will get smaller. If the first thing has an opposite effect on the second thing, the connecting line is interrupted as an open circle. Positive and Negative feedback circular systems can work side by side. Suppose an area has an abundance of good jobs then people will move in making it easier for others to follow; the increased flux of labor saturates the market and work prospect don't look as good; the work opportunities decline and people move out of the area. 20. Emotion is the major source of variability on a special memory surface. Once the emotional aspect has taken charge, no increased amount of information will take over charge. The original pattern allows the channeling of emotion. Emotion in its broadest sense provides the sole mechanism of adaptation whereby more useful patterns may gain dominance. Emotion provides substance of self and individuality. Logical thinking would be impossible without emotion. Emotion is essential to information processing. A feeling is followed by rationalization and may be just as useful as sequential approach. Much of the surface memory information is internal patterns representing the needs and emotions of the body. Attention follows the area of activation or in other words the contours of the memory surface. 21. The area of activation on the memory surface is strictly limited and cannot exceed the given size. The limited attention area settles on the most active part of the memory surface or the part most frequently used. The individual is paying attention to one part of the total. The limited attention span has an advantage to the mind: a. much is left out and discarded as irrelevant. B. Something is selected and the ability to select is important. Selection means an emotional preference and the ability to act by choice. 22. Breaking things into fragments has an advantage. Fragments have mobility. Language consists of mobile fragments that can be strung together in different ways (mathematics, science, and measurement) are fragment processes. 23. The internally produced patterns of Pain and pleasure intrude on the pure memory surface and help to direct attention. Selection is based on usefulness and instead of familiarity. The memory surface no longer deals with information for its own sake but only in terms of its usefulness. "In terms of survival and adaptation this is essential."

Brain neuron simulation closing match the biological model

 1. 400 types of neurons. The neurons are probed extensively 2. The keys is to build profiles: morphology, electrical, and iChannels 3. 200 channels to shape the electrical behavior of the neuron. The neuron choose 10% of time to form the electrical behavior. 4. The gene expression determine the phenotype and determines the electrical behavior of the cell. 99% prediction of an actual potential. 5. What is the relationship between the digital and the analogy environment? Low frequency oscillations burst are simulated. Stimulation level 4 and produces a population burst. The neocortex will have a frequency dependent event. The simulation produced a gamma oscillations around 60 and 80 hertz, a resonance. No proof of conscientiousness. 6. The goal is to build a facility for building brain circuits. In the future we are going to see simulation of different brain diseases. Brain disease cost $1 trillion a year.

What is the iChannel for brain simulation?

 1. Blue Brain is creating a simulation based neuroscience research capability 2. Neurons and synapse are created in 3D space. 3. 90 percent of the brain is in branches 4. The brain circuit diagramming must be doing semi-automatically. 5. Blue Brain will create a cellular level model of the neural transmitters in the next round of development. 6. Synaptic currents, rise times, latency, and fits to experimental data are being simulated. 7. The software allows the user to calibrate up to 30 million synaptic connections and gene, electrical, morphology, probability, connectivity, and synaptic profiles help align the connections. 8. Hodkin and Huxley model can be simulated 9. Martinotti loops can be added. 10. Simulation of low frequency oscillations followed by 60 mhz gamma bursts can be simulated. The Emergent patterns start with stimulation in one level, a spread, and radiating up the levels, radiating down the levels. 11. Blue brain will create more biological precisions, if time and money allow. 12. The software must be capable of crawling data and building models 13. The process design is experiment and reverse-engineer existing biological structures, build and apply existing neuroscience theories, write software to create simulation, and then provide feedback on assumptions of the analysis. 14. The brain has over 200 iChannels in the brain but uses about 10-20 iChannels How does the brain create perception?  1. The brain builds a model of the universe and projects the universe around us 2. Decisions support our perceptual bubble. Decisions are the key things that keep our perceptual bubble alive. Pain-killers introduce a noise into the brain, the neurons become confused, the confuse blocks the pain. 3. To perceive you have to make decisions. “I think therefore I am”. 4. Neocortex increased in size: cope with parenthood, social interaction, and instinct, complex cognitive functions. Millions of G5 processors, brain folding occurs as space become limited. Each neocortical columns creates a note. Autistic may have super plastic columns and capable of building a symphony. A million fold increased in neocortical columns allowed for more complex processes. 5. Neurons choose very carefully the neurons they connect with. Neuron branch in million of directions and connection points are the synapses. Every neuron is different. No neuron is the same. The fabric of the neuron is not the same. The pattern of how the circuit is designed does not change; humans share the same fabric. Math simulates the activity of the neocortex. Math of how neurons collect information. Math of communication between the neurons. 6. How can we create a fabric to understand each other. Circuitry does change but the pattern of design does not. We share the same fabric. Look at the electrical patterns appearing within the neocortical column, electrical objects. If you go into the neocortical column, you will see a universe of activity. Take the brain coordinates and project them into reality.

2005 Darpa Challenge

 Sebastian Thrun, autonomous vehicles used gps frames to determine velocity and direction and trajectory and a cross check error algorithm to build a path line used for navigate the vehicle. GPS gives you a sense of where you are. Laser pulses determine distance to an objects Software created motion planning and gave a sense on how to move. Motion planning gives the car the ability to drive around vertical objects. After 2005, new 3d interpretation was developed to overcome false vertical imaging situations cause from Pitch angle 3d misinterpretation. A solution was found and it used a probabilistic error model for vertical object identification where vertical objects were avoided by the car. What you have driven over is consider flat and increases the probability of direction is probable. Machine learning eliminated false learning objects. Probability was their friend. Road finding was challenging. Guassian perception ranges to horizon helped identify the road and allowed the vehicle to drive faster with 60 meter range out. Colors, shapes, and previous terrain topography were factored in the perception of what a road is. Adaptive algorithms were created for speed control to slow down the vehicle to account for rough and treacherous terrain. Reference Link 42,000 people a year die to car accidents due to human error. People sent 1.2 hours a day commuting. Aging population need autonomous vehicles to drive them. Autonomous vehicles could close the gap between vehicles allowing for faster speeds in highly congested highways without trillions of dollars in new high infrastructure investment.

Miguel Nicolelis - Brain code - Duke

 During Parkinson the brain experiences the affects of depleted levels of dopamine; eventual the dopamine treatments become ineffective and neuron stimulation treatments are the measures of last resort. The prosthetic device produces electrical stimulation to the spinal chord. The prosthetic stimulates the sensor portion of the spinal chord. During parkinson the brain activity at the motor cortex finds these cells are more synchronous and seem to be fighting each other, almost like a mild seizure, a mild epileptic seizure, but it is not a seizure.The prothestic desynchronises these neurons. Aurora the monkey 1. Nicolelis gathers and reads the signals from the motor and sensory areas of the brain for the monkey Aurora involved in walking 2. Decodes the brain code and sends them to a bipedal robot that starts walking like a monkey 3. Aurora controls a robot arm grasping pieces of banana offered and bringing the food close enough to eat. 4. Next, Nicolelis recorded brain code while Aurora controlled a joy stick while playing a special video game for the experiment. The braincode was recorded and decoded and the brain code was downloaded into computer that control a robot arm. The robot mimcked Aurora hand and arm motions.

The Next Generation of Neural Networks (Geofrey Hinton)

 1. Creating applications that do useful things 2. People are much better than computer at recognizing patterns. 3. Can we train machines to extract many layers of features by mimicking the way the brain does it. 4. Super Vector Machine worked better than Back Error Propagation. 5. What is wrong with back-propagation: it requires label data, not enough information in labels, the brain needs 10 pow 14 connection weights in 10 pow 12 second, learning does not scale well, and neurons need to send two different types of signals. Information must come from the sensor inputs. 6. The Binary Stochastic neuron outputs either a one or zero. 7. Boltzmann Machine has a set of sensory neurons connected to a hidden layer, a restricted connectivity. 8. Restricted Boltzmann Machine: The neural networks are governed by an energy function. Each possible joint configuration of the visible and hidden units has a Hopfield energy function. The energy is determined by the weights and biases. The energy of a joint configuration of he visible and hidden units determines the probability that the network will choose that configuration. By manipulating the energies of joint configurations, we can manipulate the probabilities that the model assigns to visible vectors. This gives a very simple and very effective learning algorithm. 9. The training with two layers improves abstraction without labels. Each RBM converts its data distribution into a posterior distribution over its hidden units. Task1: Learn generative weights that can convert the posterior distribution over the hidden units back into the data. Task 2: Learn to model the posterior distribution over the hidden units. 10. The model learns to generate combinations of labels and images. To perform recognition we start with a neural state of the label units and do and up-pass from the image followed by a few iterations of the top-level associative memory. 11. Using a document find similar documents in the database. Train an auto-encoder using 30 logistic units for the code layer. Add noise vector for each training case is fixed. So we still get a deterministic gradient. Document gets hashed to a 30 bit code. Similar documents match to similar documents by hash code. Hash function gives a location. Go there and look around. A document supermarket. Better than locality sensitive hashing. Comprehension done to semantic features. 12. The Restricted Boltzman Machine provides a simple way to learn a layer of features without supervision. Many layers of representation can be learned by treating the hidden states of one RBM as the visible data for training the next RBM. This creates good generative models that can be fine-tuned. BP can be used for labels and discrimination. 13. Putting lateral interactions on the nodes help create longer-range structure. What is Deep Architecture unsupervised learning? 1. Yam Lecun specialized in Visual Perception with Deep Learning 2. How do we learn representation? How do we learn by looking at objects in the world. A good learning system should build internal representation and learn by itself, learn by a few examples. 3. Human recognition of objects is very fast, 100 milliseconds. The visual cortex is composed of 10 layers of neural networks. 4. A learning machine must be capable of solving many problems. You must have a broad number of templates for template matching. More multiple layers can be used to represent more complex functions. Deep architecture trades space for time. 5. The objective of recognition classification is to identify the object independent of lighting, orientation, and background. Supervisor learning requires a huge number of training samples. Object recognition requires hundreds of thousands of images to produce an invariant representation. 6. Deep architecture is a hierarchy of modules. It takes two inputs and output is a single number, a likelihood or energy, stack them, where the output of one molecule is the input of the other. Energy function transformed into a probably function. Input Y and Features Z and (parameters W) output E(Y,Z). Low probability to things we observe. I want to make probability of the model high. Follow the gradient descent. Push down on the energy of the training samples and push up the training energy of everything else. Local minimum around training samples. 7. Y inputs to Linear Encoder (Linear, Linear-Sigmoid-Scaling) feeds to Z features outputting to Linear Decoder Feeds into cost (reconstruction energy). You need a predictor and remove the decoder. The predictor is the encoder. If Z=0 or 1 then cost object is energy low at the local min and everywhere else the energy is high. Restricting the information content code alleviates the need to push up the energy of everything. Sparsity logistic function constraint prevents the lowering of energy of everything when new inputs are added. If the training samples are not sparse enough, you don’t get good accuracy. 8. Convolutional network filters have low error using a training samples. 9. Ideas were applied to DARPA and they built LAGR (Learning Applied to Ground Robots). Machine learning to drive robots around. Identify objects that are obstacles. Stereo vison and triangulation can be used to determine the distance of something. Stereo vision is good for 12 meters. Nearsighted vision problem provides for poor decisions. The Long Range vision finds the horizon and builds bans. The system is trained both online and offline, near to far learning. Labels are classified in a hyperbolic polar map.

Verbs, Nouns, and Neural linguistic terms

This Machine learning theory matches your claim that neural representations are common across groups of people. The scientist had people look at words. He compared the brain activity of a person looking at a word and the brain activity of the person looking at the picture matching the word. The brain activity of both was similar. Then he looked at the results of groups looking at the word and found that they group had similar results. People think the same, in general neural linguistic terms. Sensory activity of the brain was connected to verbs associated with that part of the brain. Sensory: see, hear, listen, taste, touch, smell, fear. Motor: rub, lift, manipulate, run, push, move, say, eat. Abstract: fill, open, ride, approach, near, enter, drive, wear, break, clean. For each verb has 20,000 voxels/coefficients to be analyzed in a regression model. The model solves the coefficients. The model trains of the brain activity associated with the verb. In a pool of 9 people the model was 79% accurate in recognizing the verb. Eat activates voxels in the pars opercularus (gustatory cortex). Gustatory cortex is activated during eating. Push activities voxels in the postcentral gyrus (planning motor actions). Run activities voxels in the Superior temporal sulcus (body motion) The model used 10,000 words most frequent English words and predict the most activate each brain region. The scientist looked at 72 regions of the brain. Right Opercularis: wheat, beans, fruit, meat, paxil, pie, mills, break, homework, eve, potatoes, drink. (wheat) Right Superior Posterior Temporal lobe: sticks, fingers, chicken, foot, tongue, rope, sauce, nose, breasts, neck, hand, rail. Left Anterior Cingulate: poison, lovers, galaxy, harvest, sin, hindu, rays, thai, tragedy, danger, chaos, mortality. What is Machine Learning  1. Andrew Ng says machine learning is a top topic for information technology. 2. Machine learning is being applied to a wide range of problems. Machine learning groups and classifies hypothesis from medical records highlighting probable diagnosis. Machine learning is a part of the check identification process. Machine learning looks for fraudulent electronic transactions. Machine learning has been applied to Netflix to learn from movies watched and make predictions about movies you may be interested in. Machine learning optimizes the engine performance of your vehicle. 3. Machine learning is the field of study that gives machines the ability to learn without being explicitly programmed says Author Samuels. 4. Learning theory is supervised learning and begins by telling the algorithm the right answer and the algorithm learns the relationships from the inputs leading to the right answer and is expected to produce more of the right answers. 5. Unsupervised learning is used for solving many problems and grouping information into clusters. Unsupervised learning algorithms have been used to create line detection and 3D models from the angular displacements. Unsupervised learning algorithms discover structure in data. 6. Reinforced learning is based on reward function or conditional reflexive learning. If the algorithm is right then a positive feedback is provided and if the algorithm is wrong then an inhibitory feedback is inputting into the system.