How to convert your business problems to deep learning problems and solve them?
More and more business problems nowadays are using deep learning approaches to solve problems. The following is a quick guide to implementing a simple approach to solving your business problem.
- Problem Formalization: Keep on breaking down your main problem into a set of smaller subproblems until finally, you reach a set of metrics, you really care about. Let's call these your controllable levers, which will control your business. For example, for the below industries major problems can be broken down into subproblems
Healthcare — Medical Imaging (Tumor Detection). In the realm of Healthcare with a focus on Medical Imaging for tumor detection, the primary aim is the early detection of tumors from MRI scans. This boils down to classifying specific regions of an MRI scan to determine if they are tumorous or not.
- Problem Formalization: Early detection of tumors from MRI scans.
Subproblem: Classifying regions of an MRI scan as tumor or non-tumor.
Controllable Levers: Pixels of the MRI scans. - X Tensor Dimension: Assuming grayscale images of size 256x256 pixels,
- X would be [batch_size, 256, 256, 1].
- y Tensor Dimension: For binary classification (tumor or no tumor),
- y would be [batch_size, 1]
Manufacturing industry — In the Manufacturing sector, a significant concern is quality control, often executed through visual inspections. The task here is to inspect manufactured items, like car parts, for any defects using cameras
- Problem Formalization: Inspecting manufactured parts (e.g., car parts) for defects using cameras.
- Subproblem: Classifying regions of a camera capture as defective or not.
- Controllable Levers: Pixels of the camera capture.
- X Tensor Dimension: Assuming grayscale images of size 512x512 pixels,
- X would be [batch_size, 512, 512, 1].
- y Tensor Dimension: For binary classification (defective or not),
- y would be [batch_size, 1].
2. The representation problem: Once your problem levers are identified, will them as Y, and the factors that could influence them will be X. This will serve as the foundation for your model. Then select the right representation or encoding of your Y and X spaces. For alpha fold, protein folding problem by DeepMind, the problem was represented in the following way
- Y (Outcome Variable): The 3D coordinates of each amino acid or the distances and angles between amino acids.
- X (Predictor Variables): The amino acid sequence of the protein and potential evolutionary information from multiple sequence alignments.
- Encoding: Amino acids can be represented using one-hot encoding or embeddings. The evolutionary information from sequence alignments (like those from the related proteins) can also be used as input features.
3. Loss Function Definition: Based on your specific business problem, determine a loss function that will guide the learning process of your model. For Alphafold, DeepMind designed a differentiable loss function that considers the predicted distance and angle information. This loss function evaluates how close the predicted protein structure is to the true structure.
4. Problem Typology: Associate your business problem with established problem types in deep learning like classification, regression, segmentation, etc. This allows for more efficient mapping to existing solutions. Also, each problem can be solved in many faces for example image segmentation can be solved using many approaches such as Yolo, U-nets, etc.
5. Architecture Selection: This step involves a choice. Based on complexity, data availability, deployment considerations, and other factors, pick an architecture. For instance, while Multi-Layer Perceptrons (MLPs) might suffice for simple problems, Convolutional Neural Networks (CNNs) might be needed for image-related challenges. This will also require an understanding of how you are going to structure your X for example what will be the input tensor dimension of X and Y. Are you going to use pre-trained models or will require new models to be trained from scratch?
6. Training Scheme Selection: Decide on the approach for training your model, considering factors such as batch size, learning rate, and optimization algorithms. Also on the type of training, either based on architecture your solution requires supervised, unsupervised reinforcement learning, or a combination of the three.
Once the foundational aspects of the problem are settled, attention can be directed towards operationalizing the solution, leveraging standard deep learning workflows. Building data pipelines becomes paramount; this involves structuring methods to efficiently and consistently funnel data into the model. It’s not just about quantity, but also about ensuring the quality of data, which brings us to the crucial stage of data validation. Here, the data undergoes rigorous checks to identify anomalies, missing values, or any inconsistencies that could hinder the model’s performance.
Post validation, the next natural progression is model training. During this phase, the model learns from the provided data, adjusting its internal parameters to make accurate predictions or classifications. But training isn’t a one-time event. Models often undergo several iterations of training, tweaking hyperparameters, and sometimes even revisiting the initial architectural decisions to optimize performance. Alongside, it’s essential to keep tabs on potential overfitting, ensuring the model generalizes well to new, unseen data.
