Welcome and Introduction:
Swaditya Mukherjee and Angad Singh Ghataria, the leaders of the Google Developers Student Club of MPHTME Mumbai, welcomed the audience to the final day of the GDSE MPHTME AI Summit.

Previous Speakers and Topics:
Over the past two days, the summit featured insightful presentations from Mr. Krishnayak on the lifecycle of a data science project and Mr. Devangaraj Neom on augmented reality and artificial intelligence.

Historical Context of AI:
The term “artificial intelligence” was coined at Dartmouth College in 1956, and preliminary work was done by Frank Rosenblatt. AI experienced two winters before regaining prominence in computer science.

Guest Speaker Introduction:
The final keynote speaker is a respected figure in the field of AI, often referred to as the “godfather of artificial intelligence.” He is the inventor of Boltzmann machines, deep belief networks, and time delay neural networks. His research interests include variational learning, capsule networks, and other cutting-edge deep learning implementations.

Academic and Research Background:
He completed his BA in Experimental Psychology in 1970 and his PhD in Artificial Intelligence in 1978. His research group in Toronto made groundbreaking discoveries in speech recognition and object classification. He was part of the team that achieved high accuracy on the ImageNet Challenge, alongside Alex Krzyzewski and Ilya Suskeva.

Awards and Recognition:
He is the recipient of numerous awards, including the ACM Turing Award in 2018, the David E. Romelhart Prize in 2001, and the IEEE James Clerk Maxwell Gold Medal in 2016. He pioneered the use of backpropagation for learning word embeddings, popularizing the algorithm in mainstream deep learning.

Conclusion:
The audience is introduced to the final keynote speaker, an esteemed professor emeritus of computer science from the University of Toronto.

Interaction of Evolution and Learning:
Learning can accelerate evolution even within a strictly Darwinian framework where learned traits cannot be inherited genetically. Learning organisms can facilitate the transmission of information into DNA more rapidly.

Muscle Energy Functions:
Muscles generate energy functions that simplify motor control. Creating energy functions can optimize system control more effectively than direct control.

Neural Net Information Retrieval:
Neural nets trained on multiple sets of associations may appear to forget initial learning. However, this information can be recovered without disrupting subsequent learning. This recovery is possible through methods that minimize interference with subsequent learning.

Multiple Timescales in Learning:
A common theme in these topics is the occurrence of events on multiple timescales. Learning and evolution can occur simultaneously, and neural nets can learn at different rates.

Background:
The Baldwin effect refers to a phenomenon where learning can facilitate evolution by aiding in the discovery of beneficial genetic variations. Traditionally, it was thought that evolution solely relied on genetic changes and natural selection. Dr. Geoffrey Hendon presents a simplified example to illustrate the concept.

Example of Mating Circuit:
Consider an organism with a mating circuit controlled by 20 neural switches. Each switch has two alleles: “on” and “off.” There is also a third “leave it to learning” allele that allows the organism to experiment with different switch settings during its lifetime.

Evolution of Mating Circuit:
Only one specific combination of switch settings leads to successful mating and offspring production. Without learning, finding this combination through genetic evolution would require building and testing millions of organisms. However, with learning, the organism can try different switch settings during its lifetime and increase its chances of finding the optimal combination.

Baldwin Effect in Action:
Imagine a population where half the switch settings are hardwired (genetically determined) and the other half are left to learning. As the population evolves, there is a tendency to hardwire more and more of the switch settings. After building around 30,000 organisms, the population is dominated by individuals with 15 hardwired switch settings. The remaining five switches can be learned with only about 32 learning trials, making the hardwiring of the last few switches less advantageous.

Key Points:
Learning can aid evolution by helping to discover beneficial genetic variations. Hardwiring beneficial traits in DNA can reduce the number of learning trials needed to achieve optimal performance. Organisms can conduct numerous learning trials in their lifetime, making learning a more efficient and flexible mechanism compared to building and discarding entire organisms.

Inner Loops Ease Landscape Search:
Optimization landscapes can be challenging with isolated peaks in a flat terrain. Introducing learning and inner loops can transform a sharp spike into a smooth bump, simplifying the search process.

Fast Feedback Loops Simplify Control:
Precise control of a robot arm drawing on a blackboard is difficult due to fast feedback requirements. Controlling muscle stiffnesses creates an energy function that automatically positions the chalk on the board. Physics-based feedback loops enable rapid and precise control without direct intervention.

Advantages of Inner Loops for Slow Adaptive Processes:
Inner loops can greatly assist slower adaptive processes by providing a smoother fitness landscape. The outer loop can set objectives, while the inner loop optimizes based on those objectives. Examples include military commands and neural network feature interactions.

Redundant Features and Iterative Cleanup:
Top-down activation can specify which features should be active, leaving cleanup to interactions between features. Iterative processes can refine the results to achieve the desired outcome.

Example: Soldiers Forming a Square:
Instructing each soldier precisely is inefficient and requires high-bandwidth communication. Providing general instructions and allowing soldiers to interact leads to a well-formed square with minimal communication. This principle applies to various scenarios, including military formations and neural network feature interactions.

Objective Functions:
Dr. Geoffrey Hendon presents the concept of objective functions in biological adaptation. An objective function is a set of instructions that guide the development of a system to achieve specific goals.

Adaptation at Two Timescales:
The optic tectum in frogs serves as an example of adaptation at two different timescales. The optic tectum is a part of the brain that receives visual input from the eyes and is responsible for spatial processing.

Objective Functions for Connectivity:
The connectivity between the retina and the optic tectum is determined by objective functions. These objective functions specify that nearby cells in the tectum should respond to the same retina and that nearby locations in space should be coded by nearby tectal cells. The result of optimizing these objective functions is the formation of stripes in the optic tectum.

Hardwiring vs. Objective Functions:
There are two possibilities for how the connectivity between the retina and the tectum is established. One possibility is that the DNA specifies exactly what should happen (hardwiring). The other possibility is that the DNA specifies objective functions and there is a process of development during which these objective functions are optimized.

Adaptation with an Extra Eye:
Adding an extra eye to a frog provides evidence for the role of objective functions in adaptation. When a frog is given an extra eye, the connectivity between the three eyes and the optic tectum forms appropriate stripes. This suggests that the DNA is not specifying the exact connectivity but rather is providing objective functions that guide the developmental process.

Timescales of Adaptation in Biological Systems:
Biological systems adapt at various timescales. Initially, organisms relied on pure evolution, taking a whole lifetime to learn. Evolution discovered the benefit of developmental stages, allowing for optimization during growth. Gene expression and learning contribute to development, with timescales of 20 minutes and seconds, respectively. The brain exhibits multiple adaptation loops operating at different timescales, enhancing system performance.

Forgetting and Reacquiring Associations in Neural Networks:
Training neural networks with backpropagation leads to interference between different sets of associations. Learned associations become blurry as new ones are learned.

Reactivation of Forgotten Associations:
Surprisingly, retraining on a subset of associations from a forgotten set can reactivate the other associations from that set. This phenomenon resembles language reacquisition in humans, where forgotten words can be recalled after relearning a few others.

Geometric Explanation of Association Reactivation:
In a simplified 3D geometrical representation, association reactivation can be analogous to image de-blurring.

Inverse Blur:
An analogy is drawn between unblurring images and adjusting the weights of a neural network. In image processing, a blurred image can be restored by mapping it through the inverse of the blur function. Similarly, in neural networks, the goal is to find the inverse of the “blur” caused by training on a new dataset to restore the accuracy of the network on the original dataset.

Training on Multiple Datasets:
Initially, a neural network is trained on dataset A, resulting in a set of weights W that correctly classifies examples from dataset A. Subsequently, the network is trained on dataset B, modifying the weights W. From the perspective of dataset A, the new weights introduced noise, leading to incorrect outputs for examples from dataset A.

Estimating the Inverse Noise:
The objective is to estimate the inverse of the noise introduced by training on dataset B to restore the accuracy of the network on dataset A. By learning dataset A again, the network effectively unblurs the output and recovers the correct values.

How Neural Networks Learn:
Neural networks learn by descending down a steep gradient initially, then slowly along the bottom of a ravine. As learning progresses, the gradient becomes gentler until it eventually disappears.

Learning Multiple Sets of Associations:
Learning a new set of associations can disrupt the weights learned for a previous set. Retraining on a subset of examples from the previous set can restore performance on that set without significantly affecting the new set.

Recovering Old Memories:
This phenomenon is similar to how people can recover old memories after learning a new language and then returning to the country where the old language is spoken.

Partial Retraining:
Retraining on a small fraction of the weights can also help recover old memories without interfering with new memories.

Applications:
This principle may apply to recovering from amnesia and other memory-related disorders.

Neural Networks with Dual Timescale Weights:
Neural connections in neural networks have weights that adapt at different timescales. Standard slow weights learn and decay slowly, holding long-term knowledge. Fast weights learn and decay quickly, holding temporary knowledge.

Retraining and Fast Weights:
Retraining on new examples using fast weights allows for the retrieval of old memories. Fast weights act as an overlay on slow weights, preserving long-term knowledge.

Evidence for Fast Weights: Priming
Priming effects, where exposure to a stimulus improves subsequent recognition, suggest the use of fast weights. Fast weights linking embedding vector components facilitate the retrieval of related words.

The Poincare Effect and Fast Weights:
The Poincare effect, where a solution to a problem suddenly appears after a period of time, can be explained by fast weights. Fast weights that adapt to recent experiences can block or facilitate access to certain memories.

Temporary Memory and Fast Weights:
Fast weights serve as a temporary memory, storing recent experiences for a short period. They enable the retrieval of related memories through practice on a portion of the old knowledge.

Temporary Memories:
Recurrent neural networks (RNNs), like LSTMs, store temporary knowledge in the activity vectors of hidden units. External memories can store more history, but introduce challenges in managing and retrieving information. Fast weights can temporarily store memories without forgetting recent information, allowing for more processing capacity.

Fast Weights:
Fast weights implement interactions between embedding vector components, making recently encountered vectors easier to retrieve. They can integrate multiple pieces of information into a single embedding vector. Fast weights free up neural activities for more processing, enabling recursion in neural networks.

Recursion in Neural Networks:
True recursion in neural networks involves reusing the same neurons and weights for recursive calls. An example is processing a sentence with embedded clauses, where the processing of the embedded clause is reused for the whole sentence. Fast weights allow for true recursion by temporarily storing processing information and retrieving it when needed.

Challenges with Fast Weights:
Slow retrieval of weights from memory limits the efficiency of fast weights in current neural networks. Batch processing is used to share weight matrices and speed up computations, but fast weights require different matrices for each example. This results in a significant slowdown in processing speed, making fast weights less practical in large neural networks.

Future of Fast Weights:
As hardware evolves to support in-place weights, fast weights are expected to become more widely used in neural networks. This will enable more recursive and efficient neural networks, leading to advancements in various applications.

Hardwired Solutions vs. Objective Functions:
Hardwired solutions for specific scenarios can lead to inflexibility when conditions change. Specifying objective functions allows for more flexible adaptation by optimizing desired outcomes in various circumstances.

Hierarchical Management and Objectives:
Effective managers set clear objectives rather than dictating specific actions. This approach promotes creativity and adaptability in achieving desired outcomes.

Regulation of Social Media Influence:
Social media’s potential for influencing behavior raises concerns about responsible usage. Regulation is necessary to mitigate the risks associated with social media manipulation.

Environment’s Role in Adaptation:
Adaptation is fundamentally about adjusting to an environment. Structuring the environment can facilitate adaptation and learning.

Education and Structured Environments:
Education structures the learning environment, making complex concepts more accessible. Deep learning can contribute to a better understanding of learning processes and improved educational environments.

Potential for Less Painful Learning:
With a deeper understanding of learning and improved educational environments, the learning process may become less arduous.

Defining Stiffness:
Nerve signals dictate muscle stiffness rather than length. Sending specific stiffness levels to muscles allows for precise arm movement control. If the arm lags behind, adjusting stiffness can automatically correct the trajectory. Sending stiffness signals provides extra feedback and force to correct errors in perceived arm position.

Understanding Association Importance:
Experiments used purely random associations to simplify interpretation. In reality, associations have varying levels of centrality and strangeness. Learning basic associations (e.g., birds fly) before learning exceptions (e.g., penguins don’t fly) is more efficient.

Language Evolution and Model Adaptability:
Language changes over time, with older dialects becoming redundant and replaced by newer ones. Models should ideally process newer dialects better than older ones. Dr. Hendon shares a personal anecdote about his father’s language experience in Mexico.

Dr. Hendon’s Entry into Computer Science:
Dr. Hendon never formally studied or passed exams in computer science. His journey into the field was driven by his desire to understand how the brain works. Becoming a professor of computer science deterred people from asking him about his knowledge in the subject.

Computer Simulations for Neural Models:
Dr. Geoffrey Hendon began using computer simulations for neural models in the 1970s, when computers became powerful enough for such tasks. He emphasized that he is not a computer scientist and has never written a compiler.

Fast Weights vs. LSTM:
Fast weights provide higher capacity for remembering things that happened in the recent past compared to LSTM (Long Short-Term Memory) networks. LSTMs have largely been replaced by transformers, but transformers are not as good as a neural model.

Transformers and Fast Weights:
Linear transformers can be implemented using fast weights. Researchers are currently exploring this approach to improve the performance of transformers.

The Future of Artificial Intelligence:
Dr. Hendon believes that it is difficult to predict the future of artificial intelligence, and it is better to ask young people for their perspectives.

Conclusion:
The session concluded with gratitude expressed to Dr. Hendon for his insights and to the viewers for their participation.