What is Machine Learning?:
Machine learning is an approach to program creation where examples of correct outputs for given inputs are collected and used to generate a program that can handle new cases. Machine learning algorithms can produce complex programs with millions of parameters, making them suitable for tasks where simple handwritten rules are insufficient.

Examples of Machine Learning Applications:
Recognizing patterns, such as objects in real scenes, people’s identities or expressions, and spoken words. Recognizing anomalies, such as unusual credit card transactions or sensor readings in a nuclear power plant. Predicting future events, such as stock prices, currency exchange rates, or movie preferences.

MNIST Database of Handwritten Digits:
MNIST is a publicly available database of handwritten digits commonly used for training and testing machine learning algorithms. It is a convenient and well-studied task that allows for rapid experimentation and evaluation of different methods.

Deep Neural Networks for Image Recognition:
Deep neural networks with millions of parameters can recognize thousands of different object classes in high-resolution images. Errors made by these networks can provide insights into their decision-making process and help improve their performance.

Neural Networks for Speech Recognition:
Neural networks have achieved state-of-the-art results in acoustic modeling for speech recognition systems. They can make bets on thousands of alternative fragments of phonemes, aiding in the decoding process. Neural networks are being incorporated into practical speech recognition systems.

Background:
Darla Mohammed developed a system using multiple layers of binary neurons for speech recognition, using a small database and 183 labels. Pre-training was utilized to enhance the system’s performance.

Significant Improvement:
The system achieved a 20.7% error rate on a standard benchmark, outperforming the previous best result of 24.4%.

Impact on Speech Recognition:
This breakthrough led to a shift in the way speech recognition systems are designed.

Results from Leading Speech Groups:
Microsoft: Error rate reduction from 27.4% to 18.5% with deep neural network as the acoustic model. IBM: Deep neural networks outperformed their highly tuned system with an 18.8% error rate. Google: With less training data, deep neural networks reduced the error rate from 16% to 12.3% on a large vocabulary speech recognition task.

Practical Application:
Deep neural networks are now employed in Android’s voice search feature for enhanced speech recognition.

Shifting Focus:
The presentation will now delve into the topic of real neurons in the actual brain.

Inspiration from Neurons:
Real neurons inspire the design of artificial neural networks. Studying neural networks helps us comprehend the brain’s functions and develop parallel computation styles. We focus on practical problem-solving using novel learning algorithms inspired by the brain.

Cortical Neuron Structure:
A cortical neuron consists of a cell body, axon, and dendritic tree. Synapses are structures where axons from one neuron contact dendritic trees of another neuron.

Spike Generation and Synaptic Transmission:
Depolarization of the axon hillock triggers spike generation. Spikes travel along the axon as waves of depolarization. Synapses contain vesicles of transmitter chemicals that are released upon spike arrival. Transmitter molecules diffuse across the synaptic cleft and bind to receptor molecules, creating holes in the postsynaptic neuron’s membrane.

Synaptic Adaptation:
Synapses adapt by varying vesicle release or receptor sensitivity. Adaptation is a crucial aspect of learning, allowing synapses to change their effectiveness.

Advantages of Synapses:
Synapses are small, low-power, and capable of adaptation. They use locally available signals to adjust their strengths, facilitating learning.

The Challenge of Learning Rules:
Determining the rules for synaptic strength changes is a key challenge in understanding how the brain learns to perform complex computations.

Neurons and Communication:
The brain consists of neurons that receive inputs from other neurons or receptors. These neurons communicate with each other by sending spikes of activity. Synaptic weights control the effect of an input line on a neuron and can be positive or negative.

Weight Adaptation and Learning:
By adapting synaptic weights, the network learns to perform various tasks like object recognition, language comprehension, planning, and body movement control.

Storage Capacity:
The brain has approximately 10^11 neurons, each with about 10^4 weights, resulting in 10^15 or 10^14 synaptic weights. A significant portion of these weights can affect the ongoing computation in a few milliseconds, providing immense bandwidth to stored knowledge.

Modularity and Functional Localization:
The cortex is modular, meaning different parts of it learn to perform different functions. Inputs from senses go to specific parts of the cortex, influencing their functionality. Local damage to the brain affects specific functions, such as language comprehension or object recognition.

Brain Scanning and Function Location:
Brain scanners can visualize blood flow, indicating which parts of the brain are active during specific tasks. The cortex appears largely similar throughout, suggesting a universal learning algorithm.

Functional Relocation and Adaptability:
Early brain damage leads to functional relocation to other parts of the brain, demonstrating flexibility in function assignment. Experiments with baby ferrets show that rerouting sensory inputs can cause the brain to adapt and learn new functions.

General-Purpose Cortex and Flexibility:
The cortex can transform into special-purpose hardware for specific tasks based on experience. This combination of rapid parallel computation and flexibility allows for learning new functions.

Comparison with Conventional Computers:
Conventional computers use stored sequential programs for flexibility, requiring fast central processes. The brain’s flexibility comes from adapting synaptic weights, enabling parallel computation and efficient learning.

Understanding Neural Idealization:
Idealization simplifies complex systems to make them more manageable and understandable, removing non-essential details. It allows the application of mathematics, analogies to familiar systems, and subsequent refinement for increased realism. Understanding models known to be wrong can still be valuable, as long as their limitations are acknowledged.

Linear Neurons:
The output (y) is a function of bias (b), weighted activities from input lines, and synaptic weights. Graphically represented, the input-output curve is a straight line passing through zero.

Binary Threshold Neurons:
Introduced by McCulloch and Pitts, influencing von Neumann’s design of a universal computer. Computes a weighted sum of inputs and sends a spike of activity if it exceeds a threshold. McCulloch and Pitts viewed spikes as truth values of propositions, influencing the mind-as-logic paradigm. Now, the brain is seen as combining various sources of unreliable evidence, making logic less relevant.

Rectified Linear Neurons (ReLUs):
Combines properties of linear and binary threshold neurons. Computes a linear weighted sum of inputs and applies a non-linear function to the result. The input-output curve is linear above zero and zero otherwise, enabling desirable properties of linear systems while allowing for hard decisions.

Logistic Neurons:
Sigmoid neurons are commonly used in artificial neural networks. They produce a real-valued output that is a smooth and bounded function of their total input. The logistic function is used to compute the output, which ranges from 0 to 1. The derivatives of the logistic function are smooth and continuous, making it suitable for learning algorithms.

Stochastic Binary Neurons:
Stochastic binary neurons use the same equations as logistic units. They compute the total input and use the logistic function to calculate a probability of producing a spike. Instead of outputting the probability, they make a probabilistic decision and output a 1 or 0. They introduce intrinsic randomness, where the timing of spike production follows a Poisson process.

Rectified Linear Units (ReLUs):
ReLUs use a different activation function, where the output is the input if it’s positive, and 0 otherwise. Similar to stochastic binary neurons, ReLUs can be used to introduce randomness in the output by determining the rate of spike production.

Example of Machine Learning:
A simple neural network is trained to recognize handwritten shapes. The network has two layers: input neurons representing pixel intensities and output neurons representing classes. Each pixel “votes” for multiple shapes, and the shape with the most votes is recognized.

Weight Display and Learning Algorithm:
Weights between input and output units are displayed in a map, with each output unit having its own map. The strength of each connection is represented by a black and white blob, with the area indicating magnitude and the color indicating sign. A learning algorithm is used to adjust the weights based on training examples. The algorithm increments weights from active pixels to the correct class and decrements weights from active pixels to the class guessed by the network. This prevents weights from growing too large and encourages the network to learn the correct patterns.

Results:
After showing the network several hundred training examples, the weights converge to patterns that resemble templates for each shape. The network effectively learns to recognize handwritten shapes.

Challenges of Simple Neural Networks for Handwritten Digit Recognition:
Geoffrey Hinton discusses the limitations of a simple neural network in recognizing handwritten digits. The network struggles to capture variations in handwritten digits due to its inability to model allowable variations by extracting features and examining arrangements. Templates for whole shapes are insufficient for this task, as demonstrated by the examples shown.

Three Types of Machine Learning:
Hinton introduces three broad groups of machine learning algorithms: supervised learning, reinforcement learning, and unsupervised learning. Supervised learning aims to predict an output given an input vector, with the goal of minimizing the discrepancy between the target output and the actual output. Reinforcement learning focuses on selecting actions or sequences of actions to maximize rewards, which may occur occasionally. Unsupervised learning seeks to discover a good internal representation of the input data.

Supervised Learning:
Supervised learning is further divided into two categories: regression and classification. Regression involves predicting a real number or a vector of real numbers as the target output, aiming to get as close as possible to the correct value. Classification involves predicting a category or label from a set of alternatives, with the simplest case being a binary classification between positive and negative.

Model Class and Parameter Adjustment:
In supervised learning, a model class is selected, which defines the mapping from an input vector to an output using numerical parameters. The goal is to adjust these parameters to make the mapping fit the supervised training data. The discrepancy between the target output and the actual output is minimized using various measures, with the squared difference being a common choice.

Reinforcement Learning:
Reinforcement learning involves selecting actions or sequences of actions based on occasional rewards. The goal is to maximize the expected sum of future rewards, often using a discount factor to prioritize immediate rewards. Reinforcement learning is challenging due to delayed rewards, limited information from sparse rewards, and the inability to learn millions of parameters.

Unsupervised Learning:
Unsupervised learning aims to discover a good internal representation of the input data. For a long time, the machine learning community largely ignored unsupervised learning, except for clustering, which is a limited form of unsupervised learning.

Definition of Unsupervised Learning:
Clustering was commonly viewed as the only form of unsupervised learning. Hard to define the aim of unsupervised learning, a major goal is creating an internal representation of input useful for subsequent learning.

Two-Stage Learning:
Separate unsupervised and supervised/reinforcement learning to avoid using rewards/punishments to set parameters for visual systems. Example: Computing distance to a surface using disparity between images from two eyes, learned without repeatedly stubbing toes.

Goals of Unsupervised Learning:
Provide compact, low-dimensional representations of high-dimensional inputs like images. Move from high-dimensional pixel representation to a representation of a few hundred degrees of freedom, equivalent to finding a manifold in the high-dimensional space. Principal components analysis (PCA) is a limited form of this, assuming a linear manifold in the high-dimensional space. Represent input in terms of learned features, such as binary features or sparse real-valued features, for economical representation. Clustering can be viewed as a very sparse code, with one feature per cluster and only one feature active for each input.