5 4 Open Problems�(that don’t require massive models to solve)

Erik Learned-Miller

UMass Amherst

4 problems (among hundreds)

• Classifier calibration
• Lifelong learning
• Developing insights into deep and large models (without deep and large models)
• Vision for robotics

​

#1: Accuracy Calibration

99% Confidence

95% Confidence

97% Confidence

Declared a Match

#1: Accuracy Calibration

99% Confidence

95% Confidence

97% Confidence

Declared a Match

Model Confidence

True percent correct on test set

Figure from https://machinelearningmastery.com/calibrated-classification-model-in-scikit-learn/

Danger zone

Calibration is particularly bad when…

• Test data not from same distribution as training data, i.e., domain shift.
• “We only use the face recognition to arrest somebody when the confidence is over 99%!”

�Users, quite reasonably, assume confidence numbers are meaningful!

#2: �Lifelong Learning

Would this be useful?

• Yes!
• people who have built great models want to distribute their models and allow people to add capacity.
• They don’t want to (and it’s impractical to) distribute training data.
• Examples: Merlin image ID and sound ID (Grant Van Horn)
• Want to add new capacity to the model.

​

​

Inspiration from humans

• Why look at human problem solving? Because it often unveils new techniques:
• One shot learning. People are really good at it. Forces consideration of:
• Transfer learning
• Pretraining
• Meta-learning
• Lifelong learning.
• Humans continue to learn completely new topics and areas, seemingly without impairing their previous abilities.
• Memory is fixed.
• There is no way to predict how many new tasks will be learned.

#3: Insight into Big Models

• People are freaking out about “emergent properties” of large models. Strong motivation to understand them better.
• Do you need to have massive data sets or train giant models to learn how to understand big models?
• No! Not necessarily.

​

#3: Insight into Big Models

• A longstanding question about feedforward (non-recurrent) models:�How can they perform what appear to be recurrent operations?
• Example: Human-style addition
• Easy to understand how an LSTM (recurrent model with memory) can do it
• Not easy to understand how a feedforward (transformer) model can do it

​

Human-style addition with an LSTM

5 2 3 8�+ 4 9 1 6�

0

Hidden state

Inputs

Output

Human-style addition with an LSTM

5 2 3 8�+ 4 9 1 6� 4

0

Hidden state

Inputs

Output

Human-style addition with an LSTM

5 2 3 8�+ 4 9 1 6� 4

0

Hidden state

Inputs

Output

1

Human-style addition with an LSTM

5 2 3 8�+ 4 9 1 6� 5 4

0

Hidden state

Inputs

Output

1

The answer! (ICLR 2023)

The answer! (ICLR 2023)

Classic result from Circuit Complexity Theory:

- A circuit that computes a function with N recurrences can be represented by a non-recurrent circuit of depth O(log N).��The authors show how Transformer models can represent�such functions like a digital circuit in O(log N) depth.

​

​

#4: Computer Vision in service of Robotics

• Alice: a roboticist, Bob: a computer vision researcher

​

• Bob, “Hi Alice! How are you doing?”
• Alice, “I wish I had a vision algorithm that could do X!”
• Bob, “That’s easy. X is a solved problem! Try my algorithm!”
• Alice, “OK!. I tried it. It doesn’t work on my robot!”
• Bob, “Well, nice to see you Alice. See you again in 10 years.”

​

Mismatched assumptions

Opportunities in vision for robotics

• Imagine you had the ability to recognize objects, actions, activities, and also the ability to navigate, pick things up and so on…. But ….
• Every 5 minutes your long term memory was erased!
• Surprisingly, few robots actually use a long term memory.
• “Request: Find me an apple!”
• Computer vision solution: Do complex reasoning about where an apple might be based on previous associative learning and complex navigation strategies: an apple might be in the kitchen so I should navigate towards where a kitchen might be.
• Alternative solution: Remember that you saw one five minutes ago (or yesterday) and where it was.

​

Summary

• There are a wide variety of problems that don’t necessarily need massive foundational models to proceed.

Thanks!