SplineCam: Exact Visualization and Characterization of Deep Network Geometry and Decision Boundaries, CVPR 2023

Paper Link | Website

Fig: Learned implicit surface by a Neural SDF with annotations for three 2D slices (Left). SplineCam visualizations for annotated 2D slices (Rest) with decision boundary (or zero level set of the output).

Authors: Ahmed Imtiaz Humayun¹, Randall Balestriero², Guha Balakrishnan¹, Richard Baraniuk¹

¹Rice University, ²Meta AI, FAIR

Abstract: Current Deep Network (DN) visualization and interpretability methods rely heavily on data space visualizations such as scoring which dimensions of the data are responsible for their associated prediction or generating new data features or samples that best match a given DN unit or representation. In this paper, we go one step further by developing the first provably exact method for computing the geometry of a DN's mapping -- including its decision boundary -- over a specified region of the data space. By leveraging the theory of Continuous Piece-Wise Linear (CPWL) spline DNs, \textbf{SplineCam} exactly computes a DN's geometry without resorting to approximations such as sampling or architecture simplification. SplineCam applies to any DN architecture based on CPWL nonlinearities, including (leaky-)ReLU, absolute value, maxout, and max-pooling and can also be applied to regression DNs such as implicit neural representations. Beyond decision boundary visualization and characterization, SplineCam enables one to compare architectures, measure generalizability and sample from the decision boundary on or off the manifold.

splinecam_2dtoy.mp4

Video: SplineCam visualizations during training of a binary classifier MLP with width 10 and depth 5. Regions are colored by the norm of their corresponding slope parameters. Notice how the function keeps changing even when the decision boundary has converged, especially away from the training data.

Examples

Examples are placed under the ./example folder. Google colabs are also provided for some.

Model	Data	Filename	Link
MLP	Two Moons, Two Circles, Two Blobs	toy2d.py
MLP	f : 2 → 1	--
MLP - Implicit Neural Representation	2D image	2d_inr.ipynb
Scattering Network	1D Signal/Audio	scattering.ipynb
MLP	randomly initialized	visualize_mlp_random.py	Link
CNN	randomly initialized	visualize_cnn_random.py	Link
VGG11	tinyimagenet-200	-

Usage

To begin, first we have to define a 2D input space region of interest that we wish to visualize. The region of interest can be a polytopal region P at the input of the network defined via a set of vertices. Since all the region finding operations are performed in 2D, we also require a projection operator that projects vectors from the input space onto P. The following shows how to wrap any torch sequential for visualization

import torch
import splinecam

## given torch network `model` and region of interest `domain`

T = splinecam.utils.get_proj_mat(domain)

model.cuda()
model.eval()
model.type(torch.float64)

print('Wrapping model with SplineCam...')
NN = splinecam.wrappers.model_wrapper(
    model,
    input_shape=model.input_shape,
    T = T,
    dtype = torch.float64,
    device = 'cuda'
)

## check .forward() and matmul operation equivalence
print('Verifying wrapped model...')
flag =  NN.verify()
print('Model.forward and matmul equivalence check', flag)
assert flag

Compute the partitions on P along with the decision boundary and plot

regions, db_edges = splinecam.compute.get_partitions_with_db(domain,T,NN)

fig,ax = plt.subplots()
splinecam.plot.plot_partition(regions,
                              alpha=0.3, edgecolor='#a70000',ax=ax,
                              color_range=[.3,.8], linewidth=.5)
for each in db_edges:
    ax.plot(each[:,0], each[:,1], c='r', linewidth=5)

Requirements

SplineCam is mostly implemented using Pytorch and Graph-tool. All the linear algebra operations are performed using Pytorch and are vectorized, therefore scalable using GPUs. The Graph-tool operations are single threaded.

torch>=1.9,<=1.12
tqdm
networkx
python-igraph>=0.10
graph-tool==2.45
livelossplot ## to keep track of training

Setup

Clone repo

git clone https://github.com/AhmedImtiazPrio/splinecam.git

Setup graph-tool and pycairo

echo "deb http://downloads.skewed.de/apt jammy main" >> /etc/apt/sources.list
apt-key adv --keyserver keyserver.ubuntu.com --recv-key 612DEFB798507F25
apt-get update
apt-get install python3-graph-tool python3-matplotlib python3-cairo

apt purge python3-cairo
apt install libcairo2-dev pkg-config python3-dev
pip install --force-reinstall pycairo
pip install zstandard

Install misc network libraries (will be removed in future release)

pip install networkx
pip install igraph>=0.10

Install misc dependencies

pip install tqdm
pip install zstandard
pip install livelossplot
pip install imageio-ffmpeg

If PyTorch is not installed

pip install --pre torch==1.12+cu116 torchvision -f https://download.pytorch.org/whl/torch_stable.html

Release Notes

The SplineCam python toolbox can wrap any given Pytorch sequential network, containing a set of supported modules. While the number of modules currently supported is not exhaustive, we will be adding support for newer Pytorch modules over time.

Currently supported Pytorch modules are:

torch.nn.modules.linear.Linear,
torch.nn.modules.Sequential, ##also used to skip layers
torch.nn.modules.BatchNorm1d,
torch.nn.modules.Conv2d,
torch.nn.modules.BatchNorm2d,
torch.nn.modules.Flatten,
torch.nn.modules.AvgPool2d,
torch.nn.modules.Dropout

SplineCam is theoretically exact for any piecewise linear activation function, e.g., LeakyReLU, Sawtooth. Currently supported activation functions are:

torch.nn.modules.activation.ReLU,
torch.nn.modules.activation.LeakyReLU,

To do

1. Convert all igraph to graph-tool
2. Add per layer visualization support. Save meta for layer index in graph for each edge.
3. Add support for periodic activation functions
4. Add support for skip connections

Citation

@inproceedings{humayun2022splinecam,
  title={SplineCam: Exact Visualization and Characterization of Deep Network Geometry and Decision Boundaries},
  author={Humayun, Ahmed Imtiaz and Balestriero, Randall and Balakrishnan, Guha and Baraniuk, Richard},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}


@inproceedings{
humayun2022exact,
title={Exact Visualization of Deep Neural Network Geometry and Decision Boundary},
author={Ahmed Imtiaz Humayun and Randall Balestriero and Richard Baraniuk},
booktitle={NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations},
year={2022},
}