We have reached the final lesson of the course. In 08-03 we mapped where the field is heading; now for the other half of the map: what does not work well yet. Deep learning's open challenges are not footnotes: they are the reason skilled work exists in this sector, because where there are unsolved problems, professionals are needed. We will look at the technical challenges research is chasing, the challenges — more mundane but more decisive — of bringing models to production in real companies, the map of career paths, and how to turn the TecnoMarket portfolio you have built into a portfolio of your own. And we will close the way a course should close: recapping the entire journey and setting out the next steps.

1. The field's open technical challenges
2. From prototype to product: the adoption challenges
3. The map of professional opportunities
4. How to build a real portfolio from TecnoMarket
5. The full journey: course recap
6. Next steps: three specialization routes
7. Farewell

None of these problems is solved; you have already brushed against all of them in small form during the course. That is the good news: the frontier challenges are big versions of things you already understand.

Models learn the distribution of their training data and degrade — sometimes silently — when the world steps outside it. You have already lived it in everyday form: the concept drift of 06-05 (buying habits change and the demand predictor ages) is exactly this problem in slow motion; the homemade photos of 08-01 (the classifier trained on studio photos) are its instant version. Research seeks models that can detect "this doesn't look like what I saw" and abstain or adapt; today, the practical answer is still the one we built: drift monitoring, frozen validation sets and human review queues.

The frontier boasts planetary datasets, but most real problems have little labeled data, and dirty at that. You know the current weapons: transfer learning (05-03), data augmentation (05-04), self-supervised learning on unlabeled data (08-03). The open problem: truly learning from few examples, the way a person needs to see one coffee maker to recognize them all.

A generative model produces plausible outputs, not true ones. You watched it being born in your own 07-02 generator: it invented product features with complete fluency, which is why we imposed mandatory human review. In LLMs (05-05, 08-03) the problem scales: text that is self-assured, grammatically impeccable and factually false. Current mitigations: RAG (anchoring generation in real documents, 08-03), tuning the model to express uncertainty, downstream verification. A fundamental solution: none exists yet — it is perhaps the most expensive open problem of the moment, because it limits the deployment of generative systems in everything that demands accuracy.

Models are fragile against inputs designed to fool them. At the conceptual level, two families:

• Evasion attacks: tiny, imperceptible perturbations that flip the prediction (the sticker that makes a classifier see something else).
• Adaptive adversaries: the TecnoMarket case is the fraud detector from 07-03 — fraudsters are not static: they probe, observe which transactions get blocked, and adapt their behavior until they look normal. Against an adversary that learns, a model trained on yesterday's fraud is always one step behind. That is why real anti-fraud is a continuous cycle (retraining on new cases, reviewing thresholds, keeping human reviewers who spot unseen patterns), not a model trained once.

In 08-01 we saw tools (saliency, SHAP) and safeguards (human review). But truly understanding why a network with millions of parameters decides what it decides remains open: today's explanations are useful approximations, not guarantees. In the meantime, the professional norm is the one we adopted: the higher the impact of the decision, the lower the model's autonomy.

The tension of 08-02/08-03 as an open problem: the frontier advances through scale, and scale costs money, energy and concentration of power. Every advance in efficiency (quantization, distillation) relieves the symptom; the open question is whether there will be paths to capability without the scale.

The sector's uncomfortable statistic: a large share of ML projects never reaches production, and not through any fault of the models. The TecnoMarket team knows this because it has walked the whole road, and its experience sums up the real obstacles:

• Real data is dirty. The course datasets (MNIST, CIFAR-10, Fashion-MNIST) came clean and labeled. TecnoMarket's real catalog has rotated photos, inconsistent categories, reviews full of spam and transactions with empty fields. In real projects, data cleaning and preparation eats most of the time — it is where courses tend to go quiet and where professionals earn their salary.
• The model is 20%; the engineering is 80%. Look at what surrounded each model in the course: the deployment API and frozen validation (06-05), drift monitoring (06-05), the human review queue (03-04), the business thresholds in euros (07-03), the ethics audit (08-01). All of that is engineering and process, not model.fit(). The model file is the small piece of a large system.
• Maintenance never ends. A deployed model is a permanent commitment: drift, retraining, audits, dependencies that get updated, people who rotate out. "Technical debt in ML" has treacherous forms of its own: hidden dependencies between data and model (you change how a column is computed and the model degrades without warning), experiment pipelines promoted to production without refactoring, magic thresholds nobody remembers why they equal 0.7 (that is why we documented the 07-03 percentile and its matrix in euros: so the why survives the who).
• Success is measured in business terms, not in accuracy. The lesson of the confusion matrix in euros from 07-03: a model that is "worse" on metrics can be better in money, and vice versa. The projects that thrive are the ones that translate their metrics into the language of whoever decides.

The professional consequence is enormous: companies don't just need people who can train models; above all they need people who can take them to production and keep them there. That reorders the map of career paths, which is the next point.

The sector has specialized. The main roles, what they do, and which part of the course maps to each:

Honest notes about the map:

• The boundaries are porous: in a small team like TecnoMarket's, one person covers several roles; in a large one, each role is a department.
• The role with the most unmet demand is usually the one in the middle: profiles who understand the model and the engineering and the business. This course has deliberately trained you at that intersection.
• You do not need to be a researcher: the vast majority of jobs in this sector are about applying well what already exists, not inventing architectures.

You have five working projects (module 7). As they stand, they are course exercises; with directed work, they are a portfolio that opens interview doors. The difference lies in three moves:

1. Make them yours by changing the data. An interviewer recognizes CIFAR-10 and Fashion-MNIST from ten meters away. Replace them with real public datasets from a domain that interests you (open data repositories: government data portals, Kaggle, Hugging Face Datasets, UCI). Direct ideas per project:

2. Present them like a professional. Pick up the environment and tooling habits from 06-04: one clean repository per project, with a README explaining the problem, the data, the decisions (and the failures!), the metrics against a baseline, and how to reproduce it (pinned dependencies, seeds, instructions). A messy notebook with no README says "student"; a reproducible repo with justified decisions says "teammate".

3. Add the layer almost nobody adds. Pick one or two projects and take them as far as TecnoMarket went: an inference API (06-05), a small subgroup audit (08-01), an error-cost analysis (07-03). That layer — system, not just model — is exactly what the adoption section says is scarce, and it sets your portfolio apart from the remaining 90%.

Now look back. This is what you have covered, module by module, with TecnoMarket as the thread:

• Module 1 — What deep learning is, its history (from the artificial neuron to the AI winter and on to transformers), its applications, and your working environment ready. TecnoMarket was just an idea: a store with untapped data.
• Module 2 — The real fundamentals: the perceptron, activations, forward and backward propagation, losses and optimizers, and your first complete network recognizing MNIST digits. You learned the mechanism everything else reuses.
• Module 3 — CNNs: convolutions, pooling, classic architectures, and the course's first mature engineering decision: the confidence threshold with a human review queue, which stayed with us all the way to the ethics checklist.
• Module 4 — RNNs and LSTMs: review sentiment with routing, and demand prediction measured against honest baselines — the "compare yourself with something simple" discipline that reappeared in every project.
• Module 5 — The advanced techniques: GANs (and the ethical question we left planted), autoencoders and their representations, transfer learning, regularization, and the attention and transformers that closed the historical arc opened in 01-02.
• Module 6 — From the notebook to the world: TensorFlow and PyTorch, reproducible environments, and the complete deployment — API, validation on a frozen set, drift monitoring, checklist.
• Module 7 — The portfolio's five projects: the CIFAR-10 classifier (~82-85%), the description generator with temperature sampling and human review, the fraud detector with its confusion matrix in euros and its three-way decision, the Fashion-MNIST DCGAN with its handcrafted training loop, and the MobileNetV2 fine-tuning (~91%) that beat your own CNN and taught you the strategic lesson of transfer learning.
• Module 8 — The complete professional dimension: auditing bias and governing synthetic content (08-01), understanding the impact on employment, the economy and the environment (08-02), reading trends with judgment (08-03), and this final map of challenges and opportunities.

Put another way: you started not knowing what an artificial neuron was, and you finish knowing how to build, evaluate, deploy, audit and defend before a committee a complete deep learning system. That is not "having done a course"; it is a starting professional competence.

From here, the path forks according to what you enjoyed most. Three clear routes, each with its typical next topics:

Route 1: Computer vision (if you enjoyed modules 3 and 5 and projects 07-01/07-04/07-05)

• Object detection and segmentation (locating and outlining, not just classifying).
• Diffusion models in depth and vision with transformers.
• Edge AI: models on phones and devices (the MobileNet arc to the end).
• Typical courses/topics: advanced computer vision, image processing, on-device deployment.

Route 2: NLP and LLMs (if you enjoyed module 4, 05-05 and project 07-02)

• Transformers in depth: tokenization, embeddings, fine-tuning open language models.
• Engineering applications with LLMs: RAG, agents, evaluation of generative systems.
• This lesson's challenges on the front line: hallucinations and factual evaluation.
• Typical courses/topics: natural language processing, applied LLMs, information retrieval.

Route 3: MLOps and ML engineering (if you enjoyed module 6 and the "80% engineering")

• Reproducible data and training pipelines; data and model versioning.
• Monitoring, model testing, continuous deployment and deployment at scale.
• Infrastructure: containers, orchestration, inference optimization.
• Typical courses/topics: MLOps, data engineering, cloud systems architecture.

All three routes share the same methodological advice, which is this course's method: learn by building a real project end to end, with baselines, metrics and deployment, and add it to the portfolio. And whichever route you take, carry with you the three cross-cutting disciplines we practiced to the very end: always compare yourself against a baseline, keep a human in high-impact decisions, and run the ethics checklist before deploying.

• Confusing the open challenges with reasons for paralysis. That interpretability or hallucinations are unsolved does not prevent responsible deployment: it means deploying with the safeguards you learned (human review, monitoring, impact-based thresholds).
• Showing up to interviews with the course datasets as they are. MNIST and CIFAR-10 prove you followed a course; a real dataset with documented decisions proves you know how to work.
• Choosing a specialization by salary instead of affinity. All three paths are well paid and will remain so; the persistence specializing demands only comes from working on what you enjoy. Use the roles table: which modules you enjoyed is the best predictor.
• Stopping studying when the course ends. This field renews itself every few years (you saw it in 08-03). The small, sustained routine — a couple of hours a week, always one project in progress — is worth more than an annual binge.
• Underestimating what you already know. Impostor syndrome is endemic in this field because there is always something new you don't know. Objective reference: you are able to build, deploy and audit the five systems of module 7. Many working professionals have never done all five.
• Final methodological tip: when in doubt between studying something new or going deeper into what you have, almost always choose to go deeper. One project taken all the way to production teaches more than three barely started.

Six months after deployment, the recall of the 07-03 fraud detector has dropped from 83% to 61% without any alarm firing: fraudsters learned to split large purchases into several small ones from new accounts. Write a short postmortem: (a) which technical challenge from this lesson materialized, and why did the drift monitoring from 06-05 not catch it in time?; (b) propose three corrective measures of different natures (data/model, process, people); (c) what does this case tell the team about the idea of a "finished model"?

A guided self-assessment, to actually do in writing: (a) rank the course modules from the one you enjoyed most to the one you enjoyed least; (b) using the roles table and the three routes, identify your most likely route and role; (c) define your first post-course project: which TecnoMarket project you will use as a base, with which real public dataset, and which "professional layer" (API, subgroup audit or cost analysis) you will add; (d) set a realistic deadline in weeks.

As an integrative close: TecnoMarket's management asks you for a one-page ML plan for next year. It must include: one new project justified by mature trends (08-03), the maintenance plan for the five existing models (this lesson: drift, retraining, audits), the human roles needed (08-02) and the ethical safeguards (08-01). Write it in 8-12 lines: it is the course's unofficial final exam.

Solution 1. (a) The robustness challenge against an adaptive adversary materialized: fraudsters observed the system's behavior and modified their pattern until it fell within "normal". The drift monitoring from 06-05 watches for changes in the input distribution, but purchase splitting produces individually normal transactions (small amounts, new accounts: nothing odd in isolation) — the anomaly lies in the aggregate pattern across transactions, which neither the model nor the monitor was looking at; moreover, true recall is only known once fraud is confirmed weeks later, so the metric always arrives late. (b) Data/model: add per-customer and per-card aggregate variables over time windows (number of purchases in 24h, sum of amounts, account age) so the split pattern becomes visible to the model, and retrain on the newly labeled cases. Process: add delayed business metrics to the monitoring (confirmed fraud per week, in euros — the 07-03 matrix as a time series) with its own alarm, plus a fixed internal "red team" cycle that tries to fool the model before the fraudsters do. People: strengthen the human review band for new accounts and guarantee a channel through which reviewers can escalate novel patterns (it was humans, not metrics, who could have spotted it earlier). (c) That a "finished model" does not exist against an adversary that learns: anti-fraud is a permanent cycle of measurement, adaptation and human vigilance — the extreme version of the general maintenance lesson.

Solution 2. There is no single solution: it is your self-assessment. Grading criteria for yourself: the chosen route must be consistent with your favorite modules (if your top pick was module 6 and you choose vision "because it sounds better", revisit the third common mistake); the defined project must fit the deadline (a fine-tuned classifier with a subgroup audit is a realistic 2-4 weeks part-time; "building an LLM" is not a first project); and the result must end up in a public, reproducible repository, because the goal of the exercise is the first brick of your own portfolio, not a plan that stays on paper.

Solution 3. A solid plan should look something like this (one version, not the only one): "New project: visual search for the catalog (multimodality, a mature trend according to our 08-03 radar), in an internal pilot for the first half of the year, built on embeddings from a pretrained model — our proven transfer learning route. Maintenance: the five current models remain under drift monitoring plus delayed business metrics; retraining triggered by threshold (policy B from 08-02) with revalidation on frozen sets and a subgroup bias re-audit in every cycle; an explicit budget of 60% of the team's time for maintenance — we assume the model is 20% of the system. People: the human review queue reviewer and AI-content editor roles are consolidated with internal training; an MLOps profile is hired. Safeguards: no high-impact decision without a human review band; all generated content labeled and reviewed; the 08-01 ethics checklist mandatory at every deployment, with compliance/legal review — this plan is not legal advice and its regulatory fit will be validated by professionals." If your version contains these four blocks with concrete, justified decisions, you have integrated the entire course into one page.

End of the road. You started this course wondering what deep learning is; you finish it knowing what a neural network does inside (modules 1-2), how CNNs see (module 3), how RNNs and LSTMs process sequences (module 4), how GANs generate, autoencoders compress and transformers attend (module 5), how all of it is trained, saved and deployed with professional tools (module 6), and you proved it by building five real systems for TecnoMarket (module 7) that you then learned to audit, contextualize and project into the future (module 8). The TecnoMarket data team we imagined in the first lesson exists now: it is you.

Open challenges remain — models that fail outside their world, generative systems that invent, adversaries that adapt, black boxes only half opened — and that is exactly what should make you glad: a solved field would need no one new. Your next move is defined: choose your route (vision, NLP/LLMs or MLOps), turn one TecnoMarket project into a project of your own with real data and a professional layer, and keep up the small routine of continued learning. Carry with you the course's three disciplines — honest baselines, humans in the important decisions, ethics before deployment — and you will have something more valuable than any particular technique: judgment.

It has been a long journey from that first neuron. Thank you for walking it to the end. Now stop reading and go train something. The field is waiting for you.