TensorFlow Federated Tutorial Session

A Google TechTalk, 2020/7/31, presented by Google Research Staff
ABSTRACT:

[TirtharajSen]

TensorFlow federated can't be used on your windows systems. can't be installed

[kanwalzahoor]

can we use multiple dataset having different attributes in federated learning?

[hangchen]

I have the same question about server optimizer. I thought FedAvg just does a simple averaging to the model weights, so how does learning rate matter here? Any directed material I can read?

[abhidoesthat]

what if we want to find the local client metrics separately, which function should be call for the MNIST example?

[DataWizard-oy]

Timestamps (Powered by Merlin AI)
00:14 - Introduction to federated learning and TensorFlow Federated workshop.
02:16 - Federated learning enables AI without centralized data collection.
06:22 - Federated learning maintains a familiar ML workflow while decentralizing data training.
08:29 - Federated learning trains models on local data while preserving privacy.
12:42 - Federated learning iterates model training between devices and servers using federated averaging.
14:48 - Federated analytics enables data analysis on decentralized data sources.
18:26 - Federated computations utilize distributed data from clients for aggregation.
19:54 - Define Python functions for federated learning in TensorFlow.
23:08 - Iterative processes in federated learning involve state updates through loops.
24:52 - Introduction to TensorFlow Federated tutorial with setup and overview of learning process.
28:46 - Loading and exploring the MNIST dataset using TensorFlow Federated.
30:48 - Exploring non-iid behavior in federated MNIST data distribution.
34:43 - Implementing data preparation for TensorFlow Federated simulation.
36:39 - Integrating Keras models for federated training in TensorFlow Federated.
40:19 - Federated learning involves client and server optimizers for model updates.
42:17 - Initializing server state and executing federated averaging process.
46:02 - Understanding overfitting through federated training and evaluation.
47:59 - Overview of federated training model performance evaluation.
52:21 - Key considerations in federated learning and user privacy.
54:16 - Exploring federated learning for text generation and model refinement.
58:25 - Loading Shakespeare character data for model training in TensorFlow Federated.
1:00:39 - Creating TensorFlow datasets for clients by filtering and formatting data.
1:04:28 - Preprocessing text for character-based sequence prediction.
1:06:40 - Define a new character-level accuracy metric for model evaluation.
1:11:35 - Setting model weights from a Keras model is essential for federated training.
1:14:43 - Utilizing random sampling for federated learning client selection.
1:19:46 - Managing sequence length and optimizer choice in TensorFlow Federated can improve model performance.
1:21:49 - Tuning client and server optimizers is crucial for effective federated learning.
1:26:06 - Discusses client-server interactions in TensorFlow Federated.
1:28:29 - Optimizing cross-device training to mitigate timezone bias.
1:32:35 - Clients must use the same optimizer in TensorFlow Federated.
1:34:51 - Discussion on device selection and data pre-processing in federated learning.
1:39:27 - Overview of reducing communication costs in federated learning.
1:41:54 - Building a TensorFlow Federated model with Keras structure.
1:47:25 - Overview of federated evaluation and TensorBoard integration.
1:49:57 - Evaluating TensorFlow Federated model performance and optimization techniques.
1:54:57 - Uniform quantization reduces tensor values to distinct integers.
1:57:23 - Applying individual encoders for model variables in TensorFlow Federated.
2:03:08 - Model training with compression enhances efficiency without sacrificing performance.
2:06:07 - Overview of TensorFlow Federated's custom compression and research exploration.
2:10:19 - Overview of encoder customization and upcoming tutorial session.
2:12:30 - Understanding custom federated learning algorithms with TensorFlow Federated.
2:15:51 - Introduction to Federated Learning Algorithms and their Components.
2:17:33 - Outline of the initialization and update process in TensorFlow Federated.
2:20:41 - Implementing client and server updates in TensorFlow Federated.
2:22:24 - Introducing the federated core for TensorFlow Federated functionality.
2:25:52 - Understanding Federated Computations in TensorFlow Federated.
2:27:39 - Understanding federated types and computations in TensorFlow Federated.
2:30:56 - Understanding federated computations in TensorFlow Federated.
2:32:47 - Implement federated averaging using TensorFlow Federated computations.
2:36:19 - Extracting model weights and preparing client update in TensorFlow Federated.
2:38:09 - Defining the server update function in TensorFlow Federated.
2:41:43 - Overview of federated learning process with server and client weights.
2:43:34 - Overview of federated algorithm initialization and iterative updates in TensorFlow Federated.
2:47:09 - Implemented federated averaging algorithm with performance evaluation.
2:49:06 - TensorFlow Federated enables sophisticated model training through flexible client updates.
2:52:39 - Exploring custom training processes in TensorFlow Federated for cross-silo federated learning.
2:54:31 - Client weighting in federated learning can enhance model averaging.
2:58:13 - Discussing client-side weight averaging and hierarchical aggregation in TensorFlow Federated.
3:00:08 - Research focuses on defending against malicious clients in federated learning.
3:04:07 - Exploring training configurations and protocol support in TensorFlow Federated.
3:06:00 - Discussion on fault tolerance in TensorFlow Federated (TFF) and weight aggregation.
3:09:59 - Understanding TensorFlow Federated (TFF) semantics in eager and graph modes.
3:11:44 - Overview of TensorFlow Federated computation context and graph management.

[khandakermamunahmed]

Great video. A quick question is, how to preprocess cifar100 dataset?

[Dy-ygwq]

does anyone know where I can find the notebook for this?