Every question you wonder about deserves a scientific answer. The answers here rest on the pieces we publish and stay open all the way to the source.
Ask · See the sources · Be convinced
Yes — and more sharply than people expect. A network that learns a new task can lose the old one almost entirely, not partially. This is called catastrophic forgetting. The reason is simple: everything a network knows lives in the same set of weights, and a new task pulls those weights toward its own purposes.
It does, and markedly. Show the same tasks in a different order and the final error rate changes: the heavier the sequence overall, and the sharper the gaps between consecutive tasks, the more the model forgets. In a continually learning system the curriculum matters as much as the algorithm.
Usually not — it makes it longer. In a simulation of a thousand people, quarantine covering seventy percent of them left the total case count unchanged and stretched the epidemic from 49 days to 58. What changed was the peak: from around fifty percent down to under twenty-five. The aim is not to end the epidemic but to keep the number of people ill at once within what hospitals can carry.
Both matter, but speed is usually underrated. Reaching the same seventy-five percent coverage twice as fast dropped the number who fell ill from 144 to 35 in a population of a thousand. A broad campaign that starts late can do worse than a narrow one that starts early.
“Understand” is too strong a word. Classic systems reduce the sound to a few coefficients and compare it against stored templates; today’s use neural networks trained on millions of hours of speech instead. What happens in both is pattern matching — not grasping meaning.
By letting time stretch. An algorithm called dynamic time warping looks for the path that joins the start of one recording to the start of the other and the end to the end, minimising the total mismatch along the way. A syllable drawn out in one recording can then be matched to a clipped one in the other.
Only within the range that fed it. In the coffee-sales example the data sat between three and nine units of shelf space; on paper the line runs to infinity and will happily answer “what if we gave it a hundred?” That answer rests on nothing. Step outside the range and you stop producing predictions and start producing wishes.
Both — the difference is not in the method but in the habits. The machine learning tradition splits the data in two: it fits the line on one part and tests it on the part it held back, because it is easy for a model to explain data it has seen. Since it trains on labelled data and produces a numerical prediction, regression is by definition supervised learning.
It depends on the question, and that is the surprise. In an archive full of ticket numbers, error messages and file names, thirty-year-old word matching beat neural meaning search: it found 74% of the pages sought, against 65%. Meaning search weakens exactly where the word itself is what matters — a ticket number has no meaning. The best result comes from combining the two.
Without a safeguard, yes. The search layer knows nothing about permissions, only about similarity; a page can match a question perfectly and still belong to a project closed to that person. Results therefore have to be filtered by the projects someone can actually reach. An assistant either inherits an organisation's permission structure or breaches it; there is no third option.
In a software-defined network the switch has no rule yet for a new flow. So the first packet is referred to the controller, which decides and writes a rule into the switch. Later packets match that rule directly and never reach the controller. In the lab handout's sample output the first reply took 2.76 milliseconds, the second 0.052.
If no loop-prevention protocol is running, a broadcast storm begins. An ARP request sent to everyone leaves every switch through all its ports, travels round the loop, comes back and multiplies on each lap. Unlike IP packets, Ethernet frames carry no lifetime counter, so nothing stops it; links fill up and pings go unanswered. The spanning tree protocol prevents this by switching off redundant links.
