Cognition, computation, and crowdsourcing Jordan Suchow Human - PowerPoint PPT Presentation

Cognition, computation, and crowdsourcing Jordan Suchow

Human ML/AI cognition

Experiment design == Algorithm design Crowdsourcing

Experiment design == Algorithm design 1 import numpy as np 2 import judicious 3 4 z2 = np.zeros(shape=(1, 16)) 5 6 for j in range(20): 7 8 noise = NOISE_LEVEL * np.random.normal(size=(N, 16)) 9 z2s = z2 + noise 10 11 ranks = judicious.rank_the ( 12 catgory="Alan Turing", 13 images=z2s, 14 ) 15 w = [WEIGHTS[ranks.index(i)] for i in range(N)] 16 z2 = z2 + (ALPHA**j) * np.dot(noise.T, w)/(N*NOISE_LEVEL)

Example project Developing efficient methods for training face-recognition abilities.

Can we do better?

What’s changed? 1. Hardware 2. Software/algorithms 3. Data → Deep neural networks trained on massive image databases using GPUs.

A “facespace”

Now we can ask… What is the best algorithm for training face recognition abilities?

Comparing many candidate algorithms: — Hamiltonian Monte Carlo, other MCMC techniques — Stochastic gradient descent; search — Caricatures, morphs, and subtle distinctions — Active learning techniques, optimal pedagogy

And crowd-based algorithms, too.

Align learned representations with psychological representations, such as: — Trustworthiness — Age — Gender dimorphism

Cognition, computation, and crowdsourcing Jordan Suchow Human ML/AI cognition