“Facing Up to Bias,” a Presentation from Perceive
0 likes75 views
The document discusses issues related to bias in face recognition systems, emphasizing how imbalanced training datasets can lead to disproportionate error penalties based on race. It critiques traditional methods, like Generative Adversarial Networks (GANs), for failing to solve these biases, and proposes a novel approach called 'remassing' that adjusts the importance of data points based on their 'time to happiness' during training. This method aims to optimize worst-case accuracy and reduce bias by treating both rare and common data points with equal importance.
![© 2021 Perceive
Discrimination is pervasive… but not the whole story
5
• Training a neural network (typically) minimizes a loss function
• Near-universal loss function: expected value – i.e., the average – of the error
• E.g., cross-entropy H(p,q) = -Ep[log q] = average over p of –log(q)
• Suppose our FR training set has 10,000 white faces and 100 black faces
• ErrorW = average error on white faces; ErrorB = average error on black faces
• Total error is proportional to 10,000 * ErrorW + 100 * ErrorB
• Yup. Average error penalizes errors on white faces 100x as much as errors on black faces!](https://image.slidesharecdn.com/ti033teigperceive-210830153058/85/Facing-Up-to-Bias-a-Presentation-from-Perceive-5-320.jpg)
“Facing Up to Bias,” a Presentation from Perceive
- 1. © 2021 Perceive Facing Up To Bias Steve Teig Perceive
- 2. © 2021 Perceive The concerning state of face recognition (FR) 2 New York Times Khan Academy
- 3. © 2021 Perceive Backlash and emerging legislation 3
- 4. © 2021 Perceive Discrimination is pervasive… 4
- 5. © 2021 Perceive Discrimination is pervasive… but not the whole story 5 • Training a neural network (typically) minimizes a loss function • Near-universal loss function: expected value – i.e., the average – of the error • E.g., cross-entropy H(p,q) = -Ep[log q] = average over p of –log(q) • Suppose our FR training set has 10,000 white faces and 100 black faces • ErrorW = average error on white faces; ErrorB = average error on black faces • Total error is proportional to 10,000 * ErrorW + 100 * ErrorB • Yup. Average error penalizes errors on white faces 100x as much as errors on black faces!
- 6. © 2021 Perceive Of course, the trained model does better on white faces! • Total error ∝ 100 * ErrorW + 1 * ErrorB • Average error penalizes errors on white faces 100x as much as errors on black faces! • Model compression makes this problem even worse • Quantize the network, sparsify the network, etc. • If the training network must jettison some information… 6
- 7. © 2021 Perceive Why “balancing” the dataset won’t fix this 7
- 8. © 2021 Perceive For experts: why (naïve) GANs won’t fix this either • GAN: Generative Adversarial Network • Generates synthetic data points that are hard to distinguish from real data points • Can’t we use GANs to add more representative, interesting examples to the dataset? • Yes, but… • Mainstream GANs optimize only “datum looks as though from the original dataset” • What if synthetic, clean-shaven faces are easier to generate than bearded ones? • What if white faces are easier to generate than black ones? • More bias 8
- 9. © 2021 Perceive How much influence should one image have? 9 vs.
- 10. © 2021 Perceive Can we enable some images to have more influence? • In today’s deep learning, each datum appears only once per epoch during training • Loss L = 1 𝑁 σ𝑑 𝑒𝑟𝑟𝑜𝑟(𝑑) → 1 𝑁 σ𝑑 𝑚𝑎𝑠𝑠 𝑑 ∗ 𝑒𝑟𝑟𝑜𝑟(𝑑) , where σ𝑑 𝑚𝑎𝑠𝑠 𝑑 = 𝑁 • Typically, mass(d) = 1 for all d → average error • What if we increase the mass of some data points vs. others? • Mr. Muttonchops gets mass k, where all other data points get mass 𝑁−𝑘 𝑁−1 • Gradient pushes k times as hard on Mr. M • Sounds reasonable, right? 10
- 11. © 2021 Perceive Nope. Making some gradients bigger is a bad plan 11 Learning rate
- 12. © 2021 Perceive A new idea: repeated selection vs. higher “learning rate” • If d’s relative mass = k , include d (once) in each of ~k minibatches of each epoch • Look at d more than once per epoch (in different local contexts) • Now, d’s learning rate is the same as others’, but… • d moves ~k times as far per epoch • Wait a minute! How should we compute the mass of each datum? • Lossd quantifies d’s distance from happiness: i.e., lossd = 0 • Lots of papers advocate Lossd as relative importance… • Gradientd quantifies d’s current velocity on the path to happiness • Lots of papers advocate Gradientd as relative importance… 12
- 13. © 2021 Perceive Why loss and gradient are poor choices for mass 13 loss A B C D gradient gradient
- 14. © 2021 Perceive A new idea: “time to happiness” • Distance = rate * time → Time = distance / rate • Td = 𝐿𝑜𝑠𝑠𝑑 𝐺𝑟𝑎𝑑𝑖𝑒𝑛𝑡𝑑 • Want every data point to achieve happiness at (roughly) the same time • Otherwise, either stop before every data point is happy, … • Or wait for eons • Make each datum’s mass be equal to its time to happiness • Datum with more “work to do” gets more time to do it • Time to happiness is a better criterion than Lossd or Gradientd 14
- 15. © 2021 Perceive Remassing is powerful • Optimizes worst-case accuracy, rather than average accuracy • No customer really cares about average accuracy, yet everybody optimizes that! • “Accuracy: Beware of Red Herrings and Black Swans” – Embedded Vision 2020 • But wait! There’s more! • Remassing can massively accelerate training • Focus optimization effort on points with the most work to do • Most data points resemble other data points: get optimized “for free”! 15
- 16. © 2021 Perceive Facing up to bias • Remassing optimizes worst-case accuracy, not average accuracy • Treats rare data points and common data points as equally important • Treats rare (explanatory) features and common features as equally important • Remassing addresses a major source of observed bias in face recognition 16
- 17. © 2021 Perceive Resources 17 2021 Embedded Vision Summit “TinyML Is Not Thinking Big Enough” (talk) Remassing based on gradient direction https://arxiv.org/pdf/1803.09050.pdf Remassing based on loss https://arxiv.org/pdf/1511.06343.pdf Perceive https://www.perceive.io