In The Latest AI Research, Deepmind Researchers Propose Steps For Scaling Implicit Neural Representations (INRs)

This Article is written as a summay by Marktechpost Staff based on the Research Paper 'Meta-Learning Sparse Compression Networks'. All Credit For This Research Goes To The Researchers of This Project. Check out the paper, reference article.

Please Don't Forget To Join Our ML Subreddit

For a while, researchers in Deep Learning have been working to reinvent the way data is represented as functions that map to an underlying continuous signal in coordinate space. It’s a persuasive alternative to the more popular multi-dimensional array representation when such processes are approximated by neural networks.

After a rigorous architecture search, implicit Neural Representations (INRs) have recently outperformed well-established compression methods like JPEG. An essential step in making such concepts scaleable is suggested in this paper: Compression is greatly enhanced by first using cutting-edge network sparsification algorithms. A method was uncovered to further improve compression and reduce the computational cost of learning INRs that allows sparsification to be applied in the inner loop of commonly used Meta-Learning algorithms.

DeepMind suggests ways to scale implicit neural representations in a new publication called “Meta-learning: Sparse Compression Networks.” The meta-learning sparse compression networks can represent many different data types, achieving state-of-the-art performance on some of these data types.


Deep Learning relies heavily on the representation of data. The usage of implicit neural representations (INRs) as an alternative to multi-dimensional arrays has emerged as an exciting alternative. INRs have been found to surpass standard compression methods like as JPEG in recent research.

Researchers from DeepMind have developed new ways to make INRs more scalable in their recent publication Meta-Learning Sparse Compression Networks. It can effectively represent many data modalities, such as pictures, manifolds, signed distance functions, 3D shapes, and scenes, with ensuing meta-learning sparse compression networks (MSCN).

Because of the trade-off between network size and approximation quality and the need for architecture search or strong inductive biases, INRs have limits despite their outstanding compression performance. Traditional compression methods, such as JPEG from 1992, had a lower fit-to-data cost, but INRs have a far greater one.


This problem is addressed by the proposed MSCNs. With the latest sparsity techniques, the team can reduce computational costs significantly by optimizing the number of parameters used in their INRs. Meta-learning methods like MAML are also being used to learn INRs representing a single signal by finetuning them from an initialization learned with fewer optimization steps.

To re-imagine the sparsity technique as finding the best network structure for a specific task, the team relies on its efficient backpropagation and specifies optimal initialization for sparse signals. The flexibility allows it to be used in various applications, including a wide range of weights, representations, groups, and gradient sparsity.

The MSCN framework was tested using data and modalities ranging from pictures and manifolds to voxels and signed distance functions and scenes to compare it to five baselines: MAML+OneShot, MAML+IMP, Dense-Narrow, and Scratch. At least one state-of-the-art technique was surpassed by the proposed MSCN in the experiments.

Meta-learning applications requiring quick inference may find their approach particularly useful. The team notes that the techniques presented in their research can also be easily integrated with some of the baselines they’ve studied. If only a sparse portion of modulations needs to be reconstructed, the team feels a latent code technique could be competitive in future studies.